硬體 / UNIFIED MEMORY

Mac Studio M3 Ultra · 256 GB

Enough capacity for many 100B-plus low-bit models, provided the architecture is supported and optimized by the selected runtime.

一手來源

Large-memory Apple desktop256 GB

01 / 實際上限

DeepSeek V4 Flash

清單中排最前的模型可能是勉強適配或需要卸載的。日常使用建議選第一個標為 餘量充足 的結果，並保持上下文適中。

02 / 預設推論服務

MLX LM

Native Apple silicon inference, experimentation and fine-tuning. Only for Apple silicon. Use an MLX-converted model and tune KV cache size for long contexts. The bundled mlx_lm.server is an OpenAI-compatible HTTP endpoint (default 127.0.0.1:8080), but upstream warns it implements only basic security checks and is not recommended for production.

開啟指南

03 / 模型範圍

代表性開放權重模型

DeepSeek

DeepSeek V4 Flash

284B13B active · Official FP4 + FP8 mixed

估算記憶體: ~176 GB
上下文: 1M native

Community REAP build

Qwen3.8 Flash Next REAP-288

176BREAP-288 · MLX 4-bit · PLE on NVMe

估算記憶體: ~40.57 GB
上下文: 7.6K recommended on measured 48 GB Mac

OpenAI

gpt-oss-120b

117B5.1B active · MXFP4

估算記憶體: ~80 GB
上下文: 128K

Llama 4 Scout

109B17B active · INT4 MoE

估算記憶體: ~78 GB
上下文: 10M advertised

Llama 3.1 70B

70BGGUF Q4

估算記憶體: ~46 GB
上下文: 128K

Alibaba Qwen

Qwen3.6 35B-A3B

35B3B active · Q4 / NVFP4 MoE

估算記憶體: ~25 GB
上下文: 262K native

注意

適配以估算的總模型記憶體為準。驅動支援、KV 快取、多模態投影層、併發以及桌面應用都可能進一步壓縮可用餘量。目前適配：DeepSeek V4 Flash (勉強適配), Qwen3.8 Flash Next REAP-288 (餘量充足), gpt-oss-120b (餘量充足), Llama 4 Scout (餘量充足), Llama 3.1 70B (餘量充足), Qwen3.6 35B-A3B (餘量充足)。