Synthèse des preuves publiques

Adaptation du modèle local pour Mac mini et Mac Studio

Un outil de planification pour les derniers représentants de poids ouverts Qwen3.8, DeepSeek V4, GLM-5.3 et Gemma 4. Commencez par des données publiques et des calculs de capacité, séparez la durée d'exécution et les preuves de travail, puis décidez si un remplacement est justifié.

Dernière vérification : 28 août 2026 · Aide indépendante, pas un conseil Apple

Conclusion rapide

Cette page compare les représentants de chaque famille qui vous intéressent actuellement. Les versions précédentes sont conservées uniquement à titre d'enregistrement de compatibilité et exclues de la matrice principale. Même si le modèle tient dans la mémoire, l'appel d'outil, le JSON, le visuel, la concurrence et le comportement persistant peuvent être inconnus.

outils de dialogue

Vérifiez si le modèle est compatible avec votre Mac.

Choisissez votre modèle ou votre configuration Mac pour connaître les budgets pondérés, la marge de mémoire, les preuves d'exécution et les prochaines étapes vers notre conseiller gratuit.

Ouvrir l'explorateur d'adéquation

Synthèse de preuves publiques, pas une référence de première partie.

Cet outil de planification combine des cartes de modèles officielles, des métadonnées de fichiers publics GGUF/QAT, de la documentation d'exécution, des évaluations publiées et des rapports communautaires clairement étiquetés. Il ne prétend pas que Keep ou Upgrade a exécuté tous les modèles, et Apple ou les fournisseurs de modèles n'approuvent pas cette page.

Télécharger CSV Télécharger le JSON · Mêmes données sources que cette page

Les téléchargements sont des instantanés faciles à citer avec les URL sources, les niveaux de preuve et la date des données. Lien vers ce Hub comme explication canonique ; les points de terminaison de téléchargement restent en dehors du plan du site.

Comparez avec une configuration Mac

Mac mini M6 · 16GB153GB/s · US$899 · prix de base

Matrice de preuves modèles

Modèles à poids ouvert les plus récents et les plus intéressants

La matrice principale donne la priorité aux versions phares actuelles et aux représentants utiles de la planification Mac. La capacité utilise la taille de l'artefact public sélectionné lorsqu'elle est disponible ; les nombres basés sur des paramètres ne sont que des estimations de planification. Développez Sources et limites sur n’importe quelle ligne pour inspecter les signaux de communauté attribuables et leurs limites.

Famille/niveau	Modèle et version	Quantification / artefact	Capacité sur Mac sélectionné	Preuve d'exécution	Confiance	Détails
QwenPetit	Qwen3.8 27BQwen3.8 · 27B dense	Q4_0 · GGUF16.1 GB · Public file metadata	Ne tient pas-8.06GB restants après réserve	Prise en charge officiellemlx2 signaux sociaux · voir les sources	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata16.1 GB · exact public file metadata Source du runtimemlx · Prise en charge officielle Preuves communautaires (2)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet. Reddit · Qwen3.8-27B on a 24GB M4 Pro Mac miniMac mini · M4 Pro · 24GB · llama.cpp b10488 · GGUF Q4_K_M / IQ4_XSQ4_K_M 17.77GB; about 11.4 tok/s decode; about 16.6GB resident with q8 KV at 32K context · medium confidenceA first-person report says the 27B model is usable on 24GB, but the memory ceiling is visible once context and KV cache are included. Limite: One user, one build, and one workload; not a Keep or Upgrade test. Tech media / blog · Qwen3.8-27B 4-bit on a 32GB M1 ProMacBook Pro · M1 Pro · 32GB · MLX-VLM · MLX 4-bit8–8.7 tok/s; weights about 16.05GB; peak memory about 18.5–21.7GB · medium confidenceA detailed local deployment log reports a stable 32GB starting point and explicitly discourages 16GB for this 27B representative. Limite: Community log with different thermals, context, and prompt mix from the product matrix. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
QwenCourant	Qwen3.8-Flash-Next 180B (6B active)Qwen3.8-Flash-Next · 180B total / 6B active	BF16 · Safetensors360.0 GB · Public file metadata	Ne tient pas-352GB restants après réserve	Rapport communautairemlx1 signal social · voir les sources	mediumVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata360.0 GB · exact public file metadata Source du runtimemlx · Rapport communautaire Preuves communautaires (1)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet. Reddit · Latest-model comparison thread (non-Apple hardware)4× DGX Spark / other non-Apple systemsUsers discuss Qwen3.8 Flash as an exploration/subagent model; no Mac tuple or reproducible throughput is provided. · low confidenceThis is a freshness and interest signal only, not evidence that Flash Next fits a Mac configuration. Limite: Cross-hardware discussion; excluded from Mac capacity conclusions. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
QwenLimite	Qwen3.8 2.4T-A95BQwen3.8 · 2400B total / 95B active	Q4_K_M · GGUF1440.0 GB · Public model source; exact file not registered	Ne tient pas-1432GB restants après réserve	Prise en charge officiellemlx	mediumVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public model source; exact file not registered1440.0 GB · parameter planning estimate Source du runtimemlx · Prise en charge officielle Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
DeepSeekCourant	DeepSeek-V4-Flash 284B-A13BDeepSeek V4 · 284B total / 13B active	MLX-4bit · MLX170.4 GB · Public model source; exact file not registered	Ne tient pas-162.4GB restants après réserve	Rapport communautairemlx7 signaux sociaux · voir les sources	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public model source; exact file not registered170.4 GB · parameter planning estimate Source du runtimemlx · Rapport communautaire Preuves communautaires (7)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet. Reddit · DeepSeek V4 Flash on an M5 Max 128GBMac · Apple M5 Max · 128GB · ds4 / SSD-streaming path · DwarfStar IQ2XXSAbout 81GB model working set and about 31.06 tok/s in the reported run · medium confidenceA direct Apple Silicon report shows a practical 128GB path for the Flash representative. Limite: Community benchmark; exact context, thermal state, and build can change the result. Tech media / blog · DeepSeek V4 Flash on an M3 Max 128GBMacBook Pro · M3 Max · 128GB · ds4 experimental forkAbout 21 tok/s and about 81GB resident in the reported run · medium confidenceAn independent technical write-up corroborates that 128GB-class Apple Silicon can run a local Flash path. Limite: The article is tied to an experimental fork and does not validate mainline Ollama or llama.cpp. Reddit · DeepSeek V4 Flash on 64GB Macs with SSD streaming64GB Apple Silicon Macs · 64GB · ds4 SSD streaming · IQ2XXS / 2-bit classReports cluster around roughly 10–15 tok/s; technically possible, but not a comfortable in-memory workflow · medium confidence64GB appears technically viable only with aggressive quantization and storage streaming, so it is kept as a constrained edge case. Limite: Mixed community reports and different Mac generations; do not read this as a 64GB recommendation. GitHub · MLX DeepSeek V4 residency growth and resource-limit crashM4 Max 128GB and other Apple Silicon reports · 128GB · MLX-LMReproduction reports deterministic resource-limit failure around 11,300 generated tokens · high confidenceThe issue is important negative evidence: a model can fit and still fail on long generations because of runtime memory behavior. Limite: Issue state and fixes can change; check the runtime version and referenced patch before relying on MLX. GitHub · MLX cache/residency failure on an M3 Ultra 512GBMac Studio · M3 Ultra 512GB · 512GB · MLX-LMReports failure around 11,456–11,488 generated tokens in a production-style run · high confidenceEven very large unified-memory systems are not automatically stable for long-context MLX runs. Limite: This is a runtime bug report, not a capacity limit; a patched release may change the result. GitHub · DeepSeek V4 Flash Q8 garbled output on Apple NEONApple Silicon CPU / NEON · llama.cpp · Q8A specific repack path produced garbled output · high confidenceThis narrows the risk to a quantization/repack/runtime combination instead of treating every llama.cpp path as equivalent. Limite: Do not generalize one broken Q8 repack to all DeepSeek V4 Flash formats. GitHub · ds4 Apple Silicon throughput tableMac Studio · M3 Ultra 512GB · 512GB · ds4 Metal/CUDA engine · Q4 classRepository table reports short/long generation figures around 78.95 / 35.50 tok/s · medium confidenceA purpose-built engine provides a useful upper-bound signal for a large-memory Mac Studio, but it is not a mainstream runtime guarantee. Limite: Single graph worker, no batching, and engine-specific optimizations; compare only directionally. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
DeepSeekLimite	DeepSeek-V4-Pro 1.6T-A49BDeepSeek V4 · 1600B total / 49B active	MLX-4bit · MLX960.0 GB · Public model source; exact file not registered	Ne tient pas-952GB restants après réserve	Inconnullama.cpp	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public model source; exact file not registered960.0 GB · parameter planning estimate Source du runtimellama.cpp · Inconnu Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
GLMCourant	GLM-5.3-Flash 320B-A18BGLM-5.3-Flash · 320B total / 18B active	MLX-4bit · MLX204.0 GB · Public file metadata	Ne tient pas-195.99GB restants après réserve	Rapport communautairemlx2 signaux sociaux · voir les sources	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata204.0 GB · exact public file metadata Source du runtimemlx · Rapport communautaire Preuves communautaires (2)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet. Hugging Face · Community GLM-5.3-Flash MLX quantization ladderMLX · 2bit-lite / 2 / 3 / 4 / 6-bit MLXPublic metadata lists about 102.43GB to 295.60GB per variant; the 4-bit root mirror is about 203.99GB · medium confidenceThe newer repository makes the precision trade-off explicit and includes a 2bit-lite path, but it does not prove loadability or speed on a particular Mac. Limite: Community conversion; full-repository size includes duplicate root and variant folders, and no trustworthy Apple first-person benchmark is attached. Reddit · GLM-5.3 Flash comparison thread (non-Apple hardware)Dual DGX Spark / other non-Apple systemsOne report describes roughly 22 tok/s on dual Spark and a slower, more verbose interaction style · low confidenceThis indicates active community comparison but cannot be translated into a Mac recommendation. Limite: Cross-hardware, anecdotal, and workload-specific. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
GLMLimite	GLM-5.3 744B-A40BGLM-5.3 · 744B total / 40B active	MLX-4bit · MLX446.4 GB · Public model source; exact file not registered	Ne tient pas-438.4GB restants après réserve	Inconnullama.cpp	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public model source; exact file not registered446.4 GB · parameter planning estimate Source du runtimellama.cpp · Inconnu Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
Google GemmaPetit	Gemma 4 E2BGemma 4 · 2.3B effective / 5.1B total · PLE	Q4_0 · QAT3.35 GB · Public file metadata	Avec marge4.65GB restants après réserve	Rapport communautairellama.cpp1 signal social · voir les sources	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata3.35 GB · exact public file metadata Source du runtimellama.cpp · Rapport communautaire Preuves communautaires (1)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet. Tech media / blog · Apple Silicon local-AI comparison including Gemma 4Mac Studio / M4 Max 128GB test systems · 128GB · llama.cppGemma 4 appears in the comparison; the article warns that bandwidth is not the sole predictor and does not publish this exact matrix tuple · low confidenceUseful context for Apple Silicon behavior, but not an exact Gemma 4 E2B fit or performance claim. Limite: Different model/quantization details and media test methodology from this product matrix. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
Google GemmaCourant	Gemma 4 12B UnifiedGemma 4 · 11.95B dense	Q4_0 · QAT6.98 GB · Public file metadata	Limite1.02GB restants après réserve	Inconnullama.cpp	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata6.98 GB · exact public file metadata Source du runtimellama.cpp · Inconnu Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit
Google GemmaLimite	Gemma 4 26B-A4BGemma 4 · 25.2B total / 3.8B active	Q4_0 · QAT14.4 GB · Public file metadata	Ne tient pas-6.44GB restants après réserve	Rapport communautairellama.cpp	highVérifié 2026-08-28	Sources et limites Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28 Public file metadata14.4 GB · exact public file metadata Source du runtimellama.cpp · Rapport communautaire Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime. Public file metadata or a parameter estimate is not a first-party benchmark. Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing. Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result. Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks. The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here. Utilisation en conseiller gratuit

Qwen · Petit

Qwen3.8 27B

Qwen3.8 · 27B dense

Ne tient pas

Quantification / artefact: Q4_0 · GGUF16.1 GB · Public file metadata
Capacité sur Mac sélectionné: -8.06GB restants après réserve
Preuve d'exécution: Prise en charge officiellemlx
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata16.1 GB · exact public file metadata
Source du runtimemlx · Prise en charge officielle
Preuves communautaires (2)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet.
Reddit · Qwen3.8-27B on a 24GB M4 Pro Mac miniMac mini · M4 Pro · 24GB · llama.cpp b10488 · GGUF Q4_K_M / IQ4_XSQ4_K_M 17.77GB; about 11.4 tok/s decode; about 16.6GB resident with q8 KV at 32K context · medium confidenceA first-person report says the 27B model is usable on 24GB, but the memory ceiling is visible once context and KV cache are included. Limite: One user, one build, and one workload; not a Keep or Upgrade test.
Tech media / blog · Qwen3.8-27B 4-bit on a 32GB M1 ProMacBook Pro · M1 Pro · 32GB · MLX-VLM · MLX 4-bit8–8.7 tok/s; weights about 16.05GB; peak memory about 18.5–21.7GB · medium confidenceA detailed local deployment log reports a stable 32GB starting point and explicitly discourages 16GB for this 27B representative. Limite: Community log with different thermals, context, and prompt mix from the product matrix.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Qwen · Courant

Qwen3.8-Flash-Next 180B (6B active)

Qwen3.8-Flash-Next · 180B total / 6B active

Ne tient pas

Quantification / artefact: BF16 · Safetensors360.0 GB · Public file metadata
Capacité sur Mac sélectionné: -352GB restants après réserve
Preuve d'exécution: Rapport communautairemlx
Confiance: mediumVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata360.0 GB · exact public file metadata
Source du runtimemlx · Rapport communautaire
Preuves communautaires (1)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet.
Reddit · Latest-model comparison thread (non-Apple hardware)4× DGX Spark / other non-Apple systemsUsers discuss Qwen3.8 Flash as an exploration/subagent model; no Mac tuple or reproducible throughput is provided. · low confidenceThis is a freshness and interest signal only, not evidence that Flash Next fits a Mac configuration. Limite: Cross-hardware discussion; excluded from Mac capacity conclusions.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Qwen · Limite

Qwen3.8 2.4T-A95B

Qwen3.8 · 2400B total / 95B active

Ne tient pas

Quantification / artefact: Q4_K_M · GGUF1440.0 GB · Public model source; exact file not registered
Capacité sur Mac sélectionné: -1432GB restants après réserve
Preuve d'exécution: Prise en charge officiellemlx
Confiance: mediumVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public model source; exact file not registered1440.0 GB · parameter planning estimate
Source du runtimemlx · Prise en charge officielle
Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

DeepSeek · Courant

DeepSeek-V4-Flash 284B-A13B

DeepSeek V4 · 284B total / 13B active

Ne tient pas

Quantification / artefact: MLX-4bit · MLX170.4 GB · Public model source; exact file not registered
Capacité sur Mac sélectionné: -162.4GB restants après réserve
Preuve d'exécution: Rapport communautairemlx
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public model source; exact file not registered170.4 GB · parameter planning estimate
Source du runtimemlx · Rapport communautaire
Preuves communautaires (7)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet.
Reddit · DeepSeek V4 Flash on an M5 Max 128GBMac · Apple M5 Max · 128GB · ds4 / SSD-streaming path · DwarfStar IQ2XXSAbout 81GB model working set and about 31.06 tok/s in the reported run · medium confidenceA direct Apple Silicon report shows a practical 128GB path for the Flash representative. Limite: Community benchmark; exact context, thermal state, and build can change the result.
Tech media / blog · DeepSeek V4 Flash on an M3 Max 128GBMacBook Pro · M3 Max · 128GB · ds4 experimental forkAbout 21 tok/s and about 81GB resident in the reported run · medium confidenceAn independent technical write-up corroborates that 128GB-class Apple Silicon can run a local Flash path. Limite: The article is tied to an experimental fork and does not validate mainline Ollama or llama.cpp.
Reddit · DeepSeek V4 Flash on 64GB Macs with SSD streaming64GB Apple Silicon Macs · 64GB · ds4 SSD streaming · IQ2XXS / 2-bit classReports cluster around roughly 10–15 tok/s; technically possible, but not a comfortable in-memory workflow · medium confidence64GB appears technically viable only with aggressive quantization and storage streaming, so it is kept as a constrained edge case. Limite: Mixed community reports and different Mac generations; do not read this as a 64GB recommendation.
GitHub · MLX DeepSeek V4 residency growth and resource-limit crashM4 Max 128GB and other Apple Silicon reports · 128GB · MLX-LMReproduction reports deterministic resource-limit failure around 11,300 generated tokens · high confidenceThe issue is important negative evidence: a model can fit and still fail on long generations because of runtime memory behavior. Limite: Issue state and fixes can change; check the runtime version and referenced patch before relying on MLX.
GitHub · MLX cache/residency failure on an M3 Ultra 512GBMac Studio · M3 Ultra 512GB · 512GB · MLX-LMReports failure around 11,456–11,488 generated tokens in a production-style run · high confidenceEven very large unified-memory systems are not automatically stable for long-context MLX runs. Limite: This is a runtime bug report, not a capacity limit; a patched release may change the result.
GitHub · DeepSeek V4 Flash Q8 garbled output on Apple NEONApple Silicon CPU / NEON · llama.cpp · Q8A specific repack path produced garbled output · high confidenceThis narrows the risk to a quantization/repack/runtime combination instead of treating every llama.cpp path as equivalent. Limite: Do not generalize one broken Q8 repack to all DeepSeek V4 Flash formats.
GitHub · ds4 Apple Silicon throughput tableMac Studio · M3 Ultra 512GB · 512GB · ds4 Metal/CUDA engine · Q4 classRepository table reports short/long generation figures around 78.95 / 35.50 tok/s · medium confidenceA purpose-built engine provides a useful upper-bound signal for a large-memory Mac Studio, but it is not a mainstream runtime guarantee. Limite: Single graph worker, no batching, and engine-specific optimizations; compare only directionally.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

DeepSeek · Limite

DeepSeek-V4-Pro 1.6T-A49B

DeepSeek V4 · 1600B total / 49B active

Ne tient pas

Quantification / artefact: MLX-4bit · MLX960.0 GB · Public model source; exact file not registered
Capacité sur Mac sélectionné: -952GB restants après réserve
Preuve d'exécution: Inconnullama.cpp
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public model source; exact file not registered960.0 GB · parameter planning estimate
Source du runtimellama.cpp · Inconnu
Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

GLM · Courant

GLM-5.3-Flash 320B-A18B

GLM-5.3-Flash · 320B total / 18B active

Ne tient pas

Quantification / artefact: MLX-4bit · MLX204.0 GB · Public file metadata
Capacité sur Mac sélectionné: -195.99GB restants après réserve
Preuve d'exécution: Rapport communautairemlx
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata204.0 GB · exact public file metadata
Source du runtimemlx · Rapport communautaire
Preuves communautaires (2)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet.
Hugging Face · Community GLM-5.3-Flash MLX quantization ladderMLX · 2bit-lite / 2 / 3 / 4 / 6-bit MLXPublic metadata lists about 102.43GB to 295.60GB per variant; the 4-bit root mirror is about 203.99GB · medium confidenceThe newer repository makes the precision trade-off explicit and includes a 2bit-lite path, but it does not prove loadability or speed on a particular Mac. Limite: Community conversion; full-repository size includes duplicate root and variant folders, and no trustworthy Apple first-person benchmark is attached.
Reddit · GLM-5.3 Flash comparison thread (non-Apple hardware)Dual DGX Spark / other non-Apple systemsOne report describes roughly 22 tok/s on dual Spark and a slower, more verbose interaction style · low confidenceThis indicates active community comparison but cannot be translated into a Mac recommendation. Limite: Cross-hardware, anecdotal, and workload-specific.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

GLM · Limite

GLM-5.3 744B-A40B

GLM-5.3 · 744B total / 40B active

Ne tient pas

Quantification / artefact: MLX-4bit · MLX446.4 GB · Public model source; exact file not registered
Capacité sur Mac sélectionné: -438.4GB restants après réserve
Preuve d'exécution: Inconnullama.cpp
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public model source; exact file not registered446.4 GB · parameter planning estimate
Source du runtimellama.cpp · Inconnu
Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Google Gemma · Petit

Gemma 4 E2B

Gemma 4 · 2.3B effective / 5.1B total · PLE

Avec marge

Quantification / artefact: Q4_0 · QAT3.35 GB · Public file metadata
Capacité sur Mac sélectionné: 4.65GB restants après réserve
Preuve d'exécution: Rapport communautairellama.cpp
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata3.35 GB · exact public file metadata
Source du runtimellama.cpp · Rapport communautaire
Preuves communautaires (1)Les rapports publics sont de simples signaux et ne transforment jamais cette page en test du projet.
Tech media / blog · Apple Silicon local-AI comparison including Gemma 4Mac Studio / M4 Max 128GB test systems · 128GB · llama.cppGemma 4 appears in the comparison; the article warns that bandwidth is not the sole predictor and does not publish this exact matrix tuple · low confidenceUseful context for Apple Silicon behavior, but not an exact Gemma 4 E2B fit or performance claim. Limite: Different model/quantization details and media test methodology from this product matrix.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Google Gemma · Courant

Gemma 4 12B Unified

Gemma 4 · 11.95B dense

Limite

Quantification / artefact: Q4_0 · QAT6.98 GB · Public file metadata
Capacité sur Mac sélectionné: 1.02GB restants après réserve
Preuve d'exécution: Inconnullama.cpp
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata6.98 GB · exact public file metadata
Source du runtimellama.cpp · Inconnu
Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Google Gemma · Limite

Gemma 4 26B-A4B

Gemma 4 · 25.2B total / 3.8B active

Ne tient pas

Quantification / artefact: Q4_0 · QAT14.4 GB · Public file metadata
Capacité sur Mac sélectionné: -6.44GB restants après réserve
Preuve d'exécution: Rapport communautairellama.cpp
Confiance: highVérifié 2026-08-28

Sources et limites

Source officielle du modèleModèle family, parameters, and release context · Vérifié 2026-08-28
Public file metadata14.4 GB · exact public file metadata
Source du runtimellama.cpp · Rapport communautaire
Aucun témoignage Mac fiable de première main n'a été recueilli lors de cette mise à jour.La capacité reste une estimation de planification fondée sur des métadonnées publiques ; l'absence de preuve ne valide pas le runtime.
Public file metadata or a parameter estimate is not a first-party benchmark.
Community artifacts and reports do not imply vendor endorsement or Keep or Upgrade testing.
Social evidence is directional: build, context, thermals, storage streaming, and prompt mix can change the result.
Context, concurrency, tool calling, JSON, vision, and sustained behavior remain separate capability checks.
The primary matrix prioritizes current flagship releases; legacy versions remain compatibility records but are intentionally omitted here.

Utilisation en conseiller gratuit

Preuve de capacité

La capacité du flux de travail est une question distincte

Les fenêtres contextuelles sont des métadonnées de modèle. La concurrence, l'appel d'outils, la fiabilité JSON, la vision et le comportement soutenu dépendent du modèle exact, de l'exécution, du client, de la pression de la mémoire et de la tâche. Une cellule positive n’est toujours pas une garantie pour votre flux de travail.

Modèle	Contexte	Concurrence	Appels d'outils	JSON	Vision	Usage soutenu
Qwen3.8 27BPetit	Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8-Flash-Next 180B (6B active)Courant	Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8 2.4T-A95BLimite	Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Flash 284B-A13BCourant	Prise en charge officiellePublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Pro 1.6T-A49BLimite	Prise en charge officiellePublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3-Flash 320B-A18BCourant	Prise en charge officiellePublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.Source	Benchmark publicPublic GLM evaluations cover agentic or structured tasks, but they do not establish JSON validity in every client.Source	Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3 744B-A40BLimite	Prise en charge officiellePublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	Prise en charge officielleThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.Source	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 E2BPetit	Prise en charge officiellePublic model metadata lists 131,072 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 12B UnifiedCourant	Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 26B-A4BLimite	Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source	Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source	InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.	InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.	InconnuVision support is not assumed from family branding or an unrelated model variant.	InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Qwen3.8 27BPetit

Contexte: Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Qwen3.8-Flash-Next 180B (6B active)Courant

Contexte: Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Qwen3.8 2.4T-A95BLimite

Contexte: Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

DeepSeek-V4-Flash 284B-A13BCourant

Contexte: Prise en charge officiellePublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

DeepSeek-V4-Pro 1.6T-A49BLimite

Contexte: Prise en charge officiellePublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

GLM-5.3-Flash 320B-A18BCourant

Contexte: Prise en charge officiellePublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.Source
JSON: Benchmark publicPublic GLM evaluations cover agentic or structured tasks, but they do not establish JSON validity in every client.Source
Vision: Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

GLM-5.3 744B-A40BLimite

Contexte: Prise en charge officiellePublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: Prise en charge officielleThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.Source
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Gemma 4 E2BPetit

Contexte: Prise en charge officiellePublic model metadata lists 131,072 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: Prise en charge officielleThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.Source
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Gemma 4 12B UnifiedCourant

Contexte: Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Gemma 4 26B-A4BLimite

Contexte: Prise en charge officiellePublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.Source
Concurrence: Inférence systèmeConcurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.Source
Appels d'outils: InconnuNo model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON: InconnuStructured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
Vision: InconnuVision support is not assumed from family branding or an unrelated model variant.
Usage soutenu: InconnuNo public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

Comment lire les étiquettes

Prise en charge officielle La documentation officielle du modèle/runtime décrit le chemin de capacité ; le comportement exact du client doit encore être vérifié.
Benchmark public Une évaluation publiée existe, mais ce n'est pas un benchmark de débit ou de flux de travail sur Mac.
Rapport communautaire Les rapports publics ou artefacts communautaires sont des signaux, pas des tests du projet ni des garanties du fournisseur.
Inférence système Inférence limitée à partir de l'architecture capacité/runtime, pas un résultat observé.
Inconnu Aucune preuve publique fiable n'a été recueillie pour cette affirmation précise.
Vérifié par le projet Seule la combinaison exacte modèle, quantification, runtime et machine du registre de tests du projet est vérifiée.

Apportez votre vraie charge de travail

Transformez les preuves publiques en une décision personnelle.

Le conseiller gratuit ajoute vos exigences actuelles en matière de Mac, de contexte, de simultanéité, de budget, de compatibilité et hors ligne. Il peut indiquer « à vérifier » lorsque les preuves publiques ne suffisent pas.

Exécutez le conseiller gratuit

FAQ

Vérifiez les limites de preuve avant d'acheter.

S'agit-il d'un benchmark Keep ou Upgrade ?

Non. C'est un outil de synthèse et de planification de preuves publiques. Seule la combinaison exacte vérifiée dans les tests de base (smoke tests) du projet est indiquée comme vérifiée.

Comment calculer la capacité de votre Mac ?

Utilisez la taille de fichier public sélectionnée si le total exact du fichier ou du fichier fractionné est disponible, sinon utilisez l'estimation de planification explicite. Le budget du modèle réserve de la place pour macOS, le runtime et le contexte.

MoE nécessite-t-il moins de mémoire s’il s’agit d’un paramètre valide ?

Non. Les poids complets stockés correspondent à l'état de capacité. Les paramètres valides décrivent uniquement la quantité de calcul par jeton.

La longueur officielle du contexte fonctionne-t-elle sur n'importe quel Mac ?

Non. La longueur du contexte correspond aux métadonnées du modèle. Après le poids, l'exécution, le système et d'autres applications, vous devez également disposer de suffisamment de mémoire sur votre Mac sélectionné.

Pouvez-vous acheter un Mac simplement en regardant le tableau ?

Après avoir affiné votre champ de planification avec cette matrice, entrez votre Mac actuel, votre contexte, votre concurrence, vos clients, votre budget, votre compatibilité et vos exigences hors ligne dans notre conseiller gratuit. La preuve publique ne constitue pas une garantie d’achat.

apporter le travail réel

Utilisez des preuves publiques pour éclairer vos décisions d’achat.

Free Advisor ajoute votre Mac actuel, le contexte, la simultanéité, le budget, la compatibilité, les exigences hors ligne et indique qu'une vérification est requise si les preuves publiques sont insuffisantes.

Démarrer un conseiller gratuit

Keep or Upgrade est un outil indépendant ; Apple ne parraine ni n'approuve cette page.