Alibaba just put its Max-tier model on Hugging Face and ModelScope. Qwen3.8-2.4T-A95B is a text-only MoE — 2.4 trillion total parameters, ~95B active per token, 512 experts with 11 firing at a time across a 92-layer hybrid attention stack. Biggest open-weight model anyone has shipped. HN put it on the front page at 395 points.
Two constraints worth knowing before you download 2.4T of weights: no multimodal input, and thinking mode is mandatory. Every response emits a chain of thought first. There’s no flag to disable it.
What you actually run it on
Weights ship in Transformers format, with day-0 support in vLLM, SGLang and TokenSpeed. Alibaba also released an FP8 version (fine-grained, block size 128) that loses almost nothing, and unsloth already has GGUF builds up. Context is 262K native, extensible to ~1M.
If you don’t have a rack, DashScope hosts it at $2 per million input tokens and $6 output, behind OpenAI-, Anthropic- and DashScope-compatible endpoints. Typical use is the obvious one: long-context agents, coding, deep research loops.
Why it matters
Closed Max-tier capability going open is the actual story — GPQA Diamond 92.6, SWE-bench Pro 67.7. Still short of Fable 5’s 80.0 there, but it’s the strongest thing you can self-host. One catch: it’s under a custom “qwen3.8-max” license, not Apache. Open weights, not open source.
You Might Also Like
- Alibabas Qwen 3 8 Qwen3 8 max Preview Claims its Second Only to Fable 5 With Zero Benchmarks Published
- Alibaba Qwen 3 5 Just Dropped and it Brought 10 Million Milk Teas With it
- Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
- Alibaba Qwen Smart Glasses g1 s1 275 ai Glasses With Swappable Batteries and a Qwen api Backend
- Pollen Robotics Reachy Mini a 299 Desktop Humanoid That Runs 1 7m Hugging Face Models
Top AI Product 
Leave a comment