公開証拠の統合

Mac miniとMac Studioのローカルモデル適合

最新のQwen3.8、DeepSeek V4、GLM-5.3、Gemma 4オープンウェイト代表を対象にした計画ツールです。公開ファイルの事実と容量計算から始め、ランタイムと作業証拠を分けてから、買い替えが妥当か判断します。

最終確認:2026年8月28日 · 独立した判断支援でありAppleの助言ではありません

要点

このページではファミリーごとに現在関心の高い代表を比較します。旧リリースは互換性記録としてのみ残し、主マトリクスから除外します。モデルがメモリに収まっても、ツール呼び出し、JSON、視覚、並行、持続動作は未確認の場合があります。

対話ツール

モデルがMacに適合するか確認。

モデルまたはMac構成を選び、重み予算、メモリ余裕、ランタイム証拠、無料Advisorへの次の手順を確認します。

適合エクスプローラーを開く
公開証拠の統合であり、第一者ベンチマークではありません。

公式モデルカード、公開GGUF/QATファイルのメタデータ、ランタイム文書、公開評価、明示したコミュニティ報告を組み合わせる計画ツールです。Keep or Upgradeが全モデルを実行したとは主張せず、Appleやモデル提供者もこのページを推奨していません。

CSVをダウンロードJSONをダウンロード · このページと同じソースデータ

ダウンロードは出典URL、証拠レベル、データ日を含む引用可能なスナップショットです。説明の正規URLはこのHubであり、ダウンロードURLはサイトマップに含めません。

Mac mini M6 · 16GB153GB/s · US$899 · 基本価格

モデル証拠マトリクス

最新で関心の高いオープンウェイトモデル

主マトリクスは現在の主要リリースとMac計画で有用な代表を優先します。容量は可能な場合に公開ファイルサイズを使い、パラメータ値は計画用の推定です。各行の「出典と制約」を開くと、帰属可能なコミュニティ信号と限界を確認できます。

ファミリー / 層モデルと版量子化 / ファイル選択Macでの容量ランタイム証拠信頼度詳細
Qwen小型Qwen3.8 27BQwen3.8 · 27B denseQ4_0 · GGUF16.1 GB · Public file metadata適合しない-8.06GB 残り公式サポートmlx2 ソーシャル信号 · 出典を参照high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata16.1 GB · exact public file metadata
  • ランタイムソースmlx · 公式サポート
  • コミュニティ証拠 (2)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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.
無料Advisorへ
Qwen主流Qwen3.8-Flash-Next 180B (6B active)Qwen3.8-Flash-Next · 180B total / 6B activeBF16 · Safetensors360.0 GB · Public file metadata適合しない-352GB 残りコミュニティ報告mlx1 ソーシャル信号 · 出典を参照medium確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata360.0 GB · exact public file metadata
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (1)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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.
無料Advisorへ
Qwen境界Qwen3.8 2.4T-A95BQwen3.8 · 2400B total / 95B activeQ4_K_M · GGUF1440.0 GB · Public model source; exact file not registered適合しない-1432GB 残り公式サポートmlxmedium確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered1440.0 GB · parameter planning estimate
  • ランタイムソースmlx · 公式サポート
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
DeepSeek主流DeepSeek-V4-Flash 284B-A13BDeepSeek V4 · 284B total / 13B activeMLX-4bit · MLX170.4 GB · Public model source; exact file not registered適合しない-162.4GB 残りコミュニティ報告mlx7 ソーシャル信号 · 出典を参照high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered170.4 GB · parameter planning estimate
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (7)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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.
無料Advisorへ
DeepSeek境界DeepSeek-V4-Pro 1.6T-A49BDeepSeek V4 · 1600B total / 49B activeMLX-4bit · MLX960.0 GB · Public model source; exact file not registered適合しない-952GB 残り不明llama.cpphigh確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered960.0 GB · parameter planning estimate
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
GLM主流GLM-5.3-Flash 320B-A18BGLM-5.3-Flash · 320B total / 18B activeMLX-4bit · MLX204.0 GB · Public file metadata適合しない-195.99GB 残りコミュニティ報告mlx2 ソーシャル信号 · 出典を参照high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata204.0 GB · exact public file metadata
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (2)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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.
無料Advisorへ
GLM境界GLM-5.3 744B-A40BGLM-5.3 · 744B total / 40B activeMLX-4bit · MLX446.4 GB · Public model source; exact file not registered適合しない-438.4GB 残り不明llama.cpphigh確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered446.4 GB · parameter planning estimate
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
Google Gemma小型Gemma 4 E2BGemma 4 · 2.3B effective / 5.1B total · PLEQ4_0 · QAT3.35 GB · Public file metadata余裕あり4.65GB 残りコミュニティ報告llama.cpp1 ソーシャル信号 · 出典を参照high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata3.35 GB · exact public file metadata
  • ランタイムソースllama.cpp · コミュニティ報告
  • コミュニティ証拠 (1)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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.
無料Advisorへ
Google Gemma主流Gemma 4 12B UnifiedGemma 4 · 11.95B denseQ4_0 · QAT6.98 GB · Public file metadata余裕が少ない1.02GB 残り不明llama.cpphigh確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata6.98 GB · exact public file metadata
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
Google Gemma境界Gemma 4 26B-A4BGemma 4 · 25.2B total / 3.8B activeQ4_0 · QAT14.4 GB · Public file metadata適合しない-6.44GB 残りコミュニティ報告llama.cpphigh確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata14.4 GB · exact public file metadata
  • ランタイムソースllama.cpp · コミュニティ報告
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
Qwen · 小型

Qwen3.8 27B

Qwen3.8 · 27B dense

適合しない
量子化 / ファイル
Q4_0 · GGUF16.1 GB · Public file metadata
選択Macでの容量
-8.06GB 残り
ランタイム証拠
公式サポートmlx
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata16.1 GB · exact public file metadata
  • ランタイムソースmlx · 公式サポート
  • コミュニティ証拠 (2)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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.
無料Advisorへ
Qwen · 主流

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

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

適合しない
量子化 / ファイル
BF16 · Safetensors360.0 GB · Public file metadata
選択Macでの容量
-352GB 残り
ランタイム証拠
コミュニティ報告mlx
信頼度
medium確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata360.0 GB · exact public file metadata
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (1)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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.
無料Advisorへ
Qwen · 境界

Qwen3.8 2.4T-A95B

Qwen3.8 · 2400B total / 95B active

適合しない
量子化 / ファイル
Q4_K_M · GGUF1440.0 GB · Public model source; exact file not registered
選択Macでの容量
-1432GB 残り
ランタイム証拠
公式サポートmlx
信頼度
medium確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered1440.0 GB · parameter planning estimate
  • ランタイムソースmlx · 公式サポート
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
DeepSeek · 主流

DeepSeek-V4-Flash 284B-A13B

DeepSeek V4 · 284B total / 13B active

適合しない
量子化 / ファイル
MLX-4bit · MLX170.4 GB · Public model source; exact file not registered
選択Macでの容量
-162.4GB 残り
ランタイム証拠
コミュニティ報告mlx
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered170.4 GB · parameter planning estimate
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (7)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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. 制約: 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.
無料Advisorへ
DeepSeek · 境界

DeepSeek-V4-Pro 1.6T-A49B

DeepSeek V4 · 1600B total / 49B active

適合しない
量子化 / ファイル
MLX-4bit · MLX960.0 GB · Public model source; exact file not registered
選択Macでの容量
-952GB 残り
ランタイム証拠
不明llama.cpp
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered960.0 GB · parameter planning estimate
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
GLM · 主流

GLM-5.3-Flash 320B-A18B

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

適合しない
量子化 / ファイル
MLX-4bit · MLX204.0 GB · Public file metadata
選択Macでの容量
-195.99GB 残り
ランタイム証拠
コミュニティ報告mlx
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata204.0 GB · exact public file metadata
  • ランタイムソースmlx · コミュニティ報告
  • コミュニティ証拠 (2)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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. 制約: 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.
無料Advisorへ
GLM · 境界

GLM-5.3 744B-A40B

GLM-5.3 · 744B total / 40B active

適合しない
量子化 / ファイル
MLX-4bit · MLX446.4 GB · Public model source; exact file not registered
選択Macでの容量
-438.4GB 残り
ランタイム証拠
不明llama.cpp
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public model source; exact file not registered446.4 GB · parameter planning estimate
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
Google Gemma · 小型

Gemma 4 E2B

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

余裕あり
量子化 / ファイル
Q4_0 · QAT3.35 GB · Public file metadata
選択Macでの容量
4.65GB 残り
ランタイム証拠
コミュニティ報告llama.cpp
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata3.35 GB · exact public file metadata
  • ランタイムソースllama.cpp · コミュニティ報告
  • コミュニティ証拠 (1)公開レポートは信号にすぎず、プロジェクトテストには格上げしません。
  • 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. 制約: 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.
無料Advisorへ
Google Gemma · 主流

Gemma 4 12B Unified

Gemma 4 · 11.95B dense

余裕が少ない
量子化 / ファイル
Q4_0 · QAT6.98 GB · Public file metadata
選択Macでの容量
1.02GB 残り
ランタイム証拠
不明llama.cpp
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata6.98 GB · exact public file metadata
  • ランタイムソースllama.cpp · 不明
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ
Google Gemma · 境界

Gemma 4 26B-A4B

Gemma 4 · 25.2B total / 3.8B active

適合しない
量子化 / ファイル
Q4_0 · QAT14.4 GB · Public file metadata
選択Macでの容量
-6.44GB 残り
ランタイム証拠
コミュニティ報告llama.cpp
信頼度
high確認 2026-08-28
出典と制約
  • モデル公式ソースモデル family, parameters, and release context · 確認 2026-08-28
  • Public file metadata14.4 GB · exact public file metadata
  • ランタイムソースllama.cpp · コミュニティ報告
  • 今回の更新では信頼できるMacの実体験レポートを取得できませんでした。容量は公開メタデータによる計画推定であり、証拠がないことでランタイム結論を強めません。
  • 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.
無料Advisorへ

能力の証拠

ワークフロー能力は別の問いです

コンテキスト長はモデルのメタデータです。同時実行、ツール呼び出し、JSON信頼性、視覚、持続動作は、正確なモデル、ランタイム、クライアント、メモリ負荷、タスクに依存します。肯定的なセルもあなたの作業を保証しません。

モデルコンテキスト同時実行ツール呼び出しJSON視覚持続動作
Qwen3.8 27B小型
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8-Flash-Next 180B (6B active)主流
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8 2.4T-A95B境界
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Flash 284B-A13B主流
公式サポートPublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Pro 1.6T-A49B境界
公式サポートPublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3-Flash 320B-A18B主流
公式サポートPublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.出典
公開ベンチマークPublic GLM evaluations cover agentic or structured tasks, but they do not establish JSON validity in every client.出典
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3 744B-A40B境界
公式サポートPublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
公式サポートThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.出典
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 E2B小型
公式サポートPublic model metadata lists 131,072 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 12B Unified主流
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 26B-A4B境界
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
不明Vision support is not assumed from family branding or an unrelated model variant.
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8 27B小型
コンテキスト
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8-Flash-Next 180B (6B active)主流
コンテキスト
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Qwen3.8 2.4T-A95B境界
コンテキスト
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe current Qwen material documents tool or agent paths. The exact model template and runtime still need a workflow check.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Flash 284B-A13B主流
コンテキスト
公式サポートPublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
DeepSeek-V4-Pro 1.6T-A49B境界
コンテキスト
公式サポートPublic model metadata lists 1,048,576 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe official DeepSeek V4 material documents tool use in the model family. The exact local template and runtime still need a workflow check.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3-Flash 320B-A18B主流
コンテキスト
公式サポートPublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.出典
JSON
公開ベンチマークPublic GLM evaluations cover agentic or structured tasks, but they do not establish JSON validity in every client.出典
視覚
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
GLM-5.3 744B-A40B境界
コンテキスト
公式サポートPublic model metadata lists 1,000,000 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
公式サポートThe official GLM-5.3 material documents function calling and agent capabilities. It is not a Mac runtime integration benchmark.出典
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 E2B小型
コンテキスト
公式サポートPublic model metadata lists 131,072 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
公式サポートThe current model documentation describes multimodal inputs for this representative; the selected Mac runtime path still needs checking.出典
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 12B Unified主流
コンテキスト
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.
Gemma 4 26B-A4B境界
コンテキスト
公式サポートPublic model metadata lists 262,144 tokens. This is not proof that the selected Mac can hold that context with the weights.出典
同時実行
システム推論Concurrency is inferred only from memory budget and runtime architecture. No apples-to-apples public concurrency benchmark is attached.出典
ツール呼び出し
不明No model-specific public tool-calling result is strong enough for a positive claim in this snapshot.
JSON
不明Structured JSON behavior depends on the exact model template, runtime parser, and client; no exact public result is registered here.
視覚
不明Vision support is not assumed from family branding or an unrelated model variant.
持続動作
不明No public, comparable sustained Apple Silicon run is registered for this model and Mac configuration.

ラベルの読み方

  • 公式サポート 公式モデル/ランタイム文書が能力経路を示しますが、クライアントの実挙動は要確認です。
  • 公開ベンチマーク 公開評価はありますが、Macスループットやワークフローのベンチマークではありません。
  • コミュニティ報告 公開レポートやコミュニティ成果物は信号であり、プロジェクトテストやベンダー保証ではありません。
  • システム推論 容量/ランタイム構成からの限定的推論で、実測結果ではありません。
  • 不明 この主張を裏付ける信頼できる公開証拠はありません。
  • プロジェクト検証済み プロジェクトのスモーク台帳にある正確なモデル、量子化、ランタイム、マシンの組み合わせだけが検証済みです。

実際の作業を持ち込む

公開証拠を個人の判断へ。

無料Advisorは現在のMac、コンテキスト、並行数、予算、互換性、オフライン要件を追加し、証拠が不足すれば「要検証」と明記します。

無料Advisorを実行

よくある質問

購入前に証拠の境界を確認。

これはKeep or Upgradeによるベンチマークですか?

いいえ。公開証拠の統合と計画ツールです。正確なプロジェクトsmokeの組だけがプロジェクト検証済みと表示されます。

Macの容量はどう計算しますか?

正確なファイルまたは分割ファイル合計があれば選択した公開ファイルサイズを使い、それ以外は明示した計画推定を使います。モデル予算はmacOS、ランタイム、コンテキスト用の余裕を確保します。

MoEの有効パラメータなら必要メモリは少ないですか?

いいえ。保存された完全な重みが容量の条件です。有効パラメータはtokenごとの計算量を説明するだけです。

公式コンテキスト長はどのMacでも動きますか?

いいえ。コンテキスト長はモデルのメタデータです。重み、ランタイム、システム、他アプリの後にも選択Macに十分なメモリが必要です。

表だけでMacを買えますか?

このマトリクスで計画範囲を絞った後、現在のMac、コンテキスト、並行数、クライアント、予算、互換性、オフライン要件を無料Advisorに入れてください。公開証拠は購入保証ではありません。

実際の作業を持ち込む

公開証拠を自分の購入判断へ。

無料Advisorは現在のMac、コンテキスト、並行数、予算、互換性、オフライン要件を追加し、公開証拠が足りなければ要検証と示します。

無料Advisorを始める

Keep or Upgradeは独立した意思決定支援ツールであり、Appleの推奨を受けていません。