LocalForge
返回指南
模型 / ALIBABA QWEN

Qwen3.8 27B

模型卡核對於 2026-09-26 · 記憶體適配為估算值；除非有符合你配置的獨立實測，否則速度未知。

Runtime support is moving quickly. Verify the current engine release and quantized checkpoint before downloading. Intel publishes an OpenVINO int4 export (hf.co/OpenVINO/Qwen3.8-27B-int4-ov), so the OpenVINO Model Server lane is available on Intel CPUs, Arc GPUs and NPUs since OVMS v2026.4.0. On Apple silicon the engine author's own 8-bit repack (nitinpanj/Qwen3.8-27B-Splash-HQ, 27 GB, MTP draft head) ran a live session on a 64 GB M5 Pro between 24.8 and 43.5 tok/s decode from a 34,605-token context out to 190,016 tokens, but it needs that author's Splash fork on branch q8: upstream Splash 1.0 rejects non-4-bit packages, so this is not a stock-runtime expectation. Re-checked 2026-09-26: the checkpoint was read again and the platform-level M5 Max cell gained its first two measurements, neither of which is a stock-engine number.

一手來源
總參數27B
估算記憶體~19.5 GBQ4 estimate
上下文262K nativeText + Image
授權條款Apache-2.0Multilingual
01 / 適合
  • coding
  • reasoning
  • multimodal agents
02 / 估算適配
Mac Studio M5 Ultra · 96 GB餘量充足MacBook Pro M5 Max · 64 GB餘量充足Framework Desktop · Ryzen AI Max+ 395 · 128 GB餘量充足Radeon AI PRO R9700 · 32 GB餘量充足NVIDIA DGX Spark · 128 GB餘量充足
03 / 部署方式

以下是執行方式範例，不是已驗證的安裝流程。容量估算對應量化檔案，官方倉庫可能是全精度權重；MLX、GGUF 與 TensorRT 需要相符的轉換檔或引擎。

vLLM進階vllm serve Qwen/Qwen3.8-27B --dtype auto
llama.cpp中等llama-server -m /path/to/verified-model.gguf --port 9931
LM Studio簡單Search “Qwen/Qwen3.8-27B” → Load a Q4/INT4 build → Start Server
MLX LM中等mlx_lm.server --model Qwen/Qwen3.8-27B
注意

Runtime support is moving quickly. Verify the current engine release and quantized checkpoint before downloading. Intel publishes an OpenVINO int4 export (hf.co/OpenVINO/Qwen3.8-27B-int4-ov), so the OpenVINO Model Server lane is available on Intel CPUs, Arc GPUs and NPUs since OVMS v2026.4.0. On Apple silicon the engine author's own 8-bit repack (nitinpanj/Qwen3.8-27B-Splash-HQ, 27 GB, MTP draft head) ran a live session on a 64 GB M5 Pro between 24.8 and 43.5 tok/s decode from a 34,605-token context out to 190,016 tokens, but it needs that author's Splash fork on branch q8: upstream Splash 1.0 rejects non-4-bit packages, so this is not a stock-runtime expectation. Re-checked 2026-09-26: the checkpoint was read again and the platform-level M5 Max cell gained its first two measurements, neither of which is a stock-engine number.