Qwen's 2.4T Qwen3.8-Max and a single-GPU 27B go open this week
Qwen3.8-Max is already live via API at $2/$6 per Mtok; the open weights hit Hugging Face and ModelScope this week, with the 27B built to run on one 4090.
Copy markdown
The drop: a 2.4T flagship goes open
Alibaba has confirmed open weights for Qwen3.8-Max (2.4T params, 95B active) and a smaller Qwen3.8-27B, both shipping this week to Hugging Face and ModelScope. As of publish neither official repo is live yet and the Max stays API-only on QwenCloud, so watch the Qwen org before trusting anything claiming to be it.
The 27B is the one you'll actually run
Qwen3.8-27B is sized for a single GPU: about 16GB VRAM at Q4_K_M, so it fits a 4090, or ~27GB for FP8 on an L40S. That makes it a genuinely local coding and agent model, unlike the Max, which needs roughly 20 H100s (~1.6TB) at full precision.
What the Max actually scores
In third-party evals the Max posts 87.3% on SWE-bench and ranks #4 on Frontend Code Arena (1,668 Elo), and Alibaba pitches it for multi-day autonomous coding. API pricing is $2 in / $6 out per Mtok ($0.25 cached), with a 1M-token context window.
Read the license before you ship
Don't assume Apache 2.0. Recent Qwen weights shipped under the Tongyi Qianwen license, which forces a commercial-terms conversation past 100M monthly active users, and no LICENSE file exists yet. Check it on the official repo before you build a product on these weights.
Build this weekend: a private 27B coding agent
Once the 27B lands you can run a fully offline coding agent on one consumer GPU: no per-token bill, no code leaving your machine. Star the Qwen GitHub and Hugging Face org now so the weight commit hits your feed the second it drops.