Qwen3.8's open weights land text-only — and the 27B still isn't out
The 2.4T-A95B MoE is downloadable now under a custom non-Apache license, but vision and the default 1M context are gone — and it's a 4.9TB datacenter job.
Copy markdown
Update: the weights are live, but text-only and thinking-only
Qwen/Qwen3.8-2.4T-A95B and an FP8 sibling are on Hugging Face now — 2.4T total, 95B active, 262K native context extensible to ~1M. But multimodal input is dropped, thinking mode can't be turned off, and the Max API's built-in tools aren't included.
The license isn't Apache — read it before you ship
Qwen 3.5 and 3.6 shipped Apache-2.0; this lands under a bespoke “qwen3.8-max” license instead. Check the commercial terms before building on it, and pull only from the official Qwen org on Hugging Face or ModelScope, not mirrors.
4.9 TB, datacenter-only — and the 27B you wanted is missing
The full checkpoint is ~4.9TB across 224 files; even the FP8 build is datacenter-scale, not a home-GPU download. The 27B model promised for the same week — the one most builders were actually waiting for — still hasn't shipped.
What to do this week
Need vision or 1M-by-default context? Stay on the Qwen3.8-Max API for now. Self-hosting? Grab the FP8 repo or unsloth's GGUF quants, but budget real hardware — and watch the Qwen org for the 27B before planning any local build.