MODEL / ALIBABA QWEN

Qwen3.6 35B-A3B

All 35B parameters must be stored even though approximately 3B are active per token.

Źródło pierwotne

Wszystkie parametry35B3B active / token

Szacowana pamięć~25 GBQ4 / NVFP4 MoE

Kontekst262K nativeText + Image

LicenseApache-2.0Multilingual

01 / NAJLEPSZY DO

fast agents
coding
tool use

02 / SZACOWANE DOPASOWANIE

Mac Studio M5 Ultra · 96 GBSwobodnie MacBook Pro M5 Max · 64 GBSwobodnie Framework Desktop · Ryzen AI Max+ 395 · 128 GBSwobodnie Radeon AI PRO R9700 · 32 GBNa styk NVIDIA DGX Spark · 128 GBSwobodnie

03 / URUCHOM PRZEZ

These are runtime examples, not validated installation recipes. Fit estimates assume quantized files; official repositories may hold full-precision weights. MLX, GGUF and TensorRT require matching conversions or engines.

vLLMAdvancedvllm serve Qwen/Qwen3.6-35B-A3B --dtype auto

llama.cppMediumllama-server -m /path/to/verified-model.gguf --port 9931

LM StudioEasySearch “Qwen/Qwen3.6-35B-A3B” → Load a Q4/INT4 build → Start Server

OllamaEasyollama run <model-tag>

Uwaga

All 35B parameters must be stored even though approximately 3B are active per token.