Qwen3.8-Max open weights land: a 2.4T-param model you can download
China's largest open model is downloadable now — but Apache 2.0 is gone, the license has strings, and only the 4.89TB flagship shipped, not the handy 27B.
Copy markdown
2.4T total, 95B active — and it's yours
The flagship is a sparse MoE: 2.4T parameters, ~95B active per token across 512 experts (11 fire at once), with BF16 and an official FP8 checkpoint on Hugging Face and ModelScope. Native 256K context stretches toward 1M. Catch: the open weights are text-only — the hosted API's vision didn't come with them.
Read the license before you build
Past Qwen models shipped Apache 2.0; this one carries a custom 'qwen3.8-max' license with a reported revenue-sharing requirement for large commercial users, and the thresholds aren't published. If you're shipping a product on it, check the terms first.
Running it locally is a real project
Community unsloth GGUFs span 1-bit (~397GB) to full BF16 (4.89TB); a 4-bit build is ~1.31TB. Nvidia's reference FP8 deployment used a GB300 NVL72 — 72 Blackwell GPUs. Translation: it's a big-RAM box or a multi-GPU rig, not your laptop.
Don't want the hardware? Just call it
The model's been live via API since Aug 3 on QwenCloud (qwen3.8-max) and Vercel's AI Gateway (alibaba/qwen3.8-max) at $2/M in, $6/M out, and $0.25/M cached, with OpenAI- and Anthropic-compatible endpoints. Together AI and HuggingChat host it too.
Where it actually lands
Vendor and community numbers put it near the frontier without topping it: 87.3% on SWE-bench (Opus 4.8: 89.2%), #4 on Frontend Code Arena at 1,668 Elo, and a Vals Index tie with Opus 4.7 at 66.1. Alibaba calls it 'second only to Fable 5' — treat that as an unverified claim.
The 27B everyone wanted is still MIA
The genuinely deployable Qwen3.8-27B was promised alongside the flagship but hasn't shipped; community trackers are watching for it around Aug 15. If 2.4T is out of reach, that's the release to wait for.