Qwen3.5 397B

About

How it compares

Of the 6 chat models Outmake recommends for a 512 GB Mac that LM Arena rates, it places 3rd in the public vote, behind GLM 5.3 and Gemma 4 31B.

In LM Arena's public vote among open models, it ranks 4th of 15 for reading pictures and 9th of 62 for text.

Outmake estimates it is 4th fastest of the 7 chat models for a 512 GB Mac. Gemma 4 26B writes about 2.8 times as fast and also reads pictures.

Good to know

Part of its training was practice in simulated worlds with up to a million agents, on tasks that grew harder.

Qwen says training on pictures cost almost no efficiency compared with training on text alone.

Alibaba Cloud runs a hosted version called Qwen3.5-Plus, with extra features for business use.

Sources: Qwen/Qwen3.5-397B-A17B model card, Outmake's speed estimates, LM Arena and Hugging Face. Every line comes from one of these.

Run It Anywhere

Qwen3.5 397B is an open model, so any app that runs GGUF files can run it. These take the UD-Q3_K_XL size from unsloth straight from Hugging Face.

• LM StudioMac, Windows, LinuxOpen
• JanMac, Windows, LinuxOpen
• MstyMac, Windows, LinuxOpen
• OllamaMac, Windows, Linuxollama run hf.co/unsloth/Qwen3.5-397B-A17B-GGUF:UD-Q3_K_XLCopy
• llama.cppMac, Windows, Linuxllama-server -hf unsloth/Qwen3.5-397B-A17B-GGUF:UD-Q3_K_XLCopy
• Docker Model RunnerMac, Windows, Linuxdocker model run hf.co/unsloth/Qwen3.5-397B-A17B-GGUF:UD-Q3_K_XLCopy