If you have been holding out for an open-weight model that can touch the frontier without buying a small data center, August 3 was the day to start paying attention. Alibaba’s Qwen team shipped Qwen3.8-Max at 2.4 trillion total parameters and told the public, via Alibaba Cloud’s official post, that the open weights are coming “next week.” Alongside it comes Qwen3.8-27B, and Daniel Han of Unsloth has already confirmed the 27B will run in 17 GB of VRAM. That number is the headline.
What landed on August 3
Qwen3.8-Max has 2.4 trillion total parameters and 95 billion active per token, and MarkTechPost describes it as a mixture-of-experts model. The Max class had previously been hosted-only; this is the first Max-class Qwen to ship open weights. The preview went live in mid-July at one-tenth the standard price through Qwen’s Token Plan, and the full release is now generally available through QwenCloud and the Model Studio API. Weights are scheduled to land on Hugging Face and ModelScope around August 10, per the Decoder’s reporting.
The Max tier is designed for long-horizon autonomy. The team’s case studies include a 16-day, 265-commit, 127-PR open-source build called oh-my-cli that ran without intervention, a five-day research-paper reproduction that beat the source paper on AIME24 by 2.7 points after roughly 125 GPU-hours of training across 33 jobs, and a chip-design pass that cut logic gates from 8,298 to 678 over about 500 iterations while shrinking die area from 106×106 µm to 46×46 µm (an 81% reduction). A separate E-Commerce-Bench run took the agent from 100,000 yuan to 416,252 yuan in revenue, 38% more than GLM 5.2 and 2.5× the predecessor Qwen3.7-Max.
For everyday use, the API pricing is $2.00 per million input tokens and $6.00 per million output tokens, with implicit cache reads at $0.25 and explicit cache creation at $2.50. The context window is a flat 1 million tokens, with a maximum output of 131K tokens and a reasoning budget of 262K. The server takes text and image input today; video is positioned for Qwen-MM-Plugins rather than as a native API input.
Where it actually lands on benchmarks
Alibaba’s own tables, cross-referenced through MarkTechPost and the Apidog head-to-head with Kimi K3, are the only data we have right now - there is no independent head-to-head run yet. On Terminal-Bench 2.1, Qwen3.8-Max scores 86.6, behind GPT-5.6 Sol’s 88.8 and slightly ahead of Claude Opus 4.8 and Fable 5, which both register 84.6. On SWE-bench Pro it lands at 67.7 versus Fable 5’s 80.0 and Claude Opus 4.8’s 69.2. PaperBench is 93.0 (highest of the named group), GPQA Diamond 92.6 (tied with Fable 5, one point under GPT-5.6 Sol), and HLE 43.6 (trailing both Anthropic models and GPT-5.6 Sol). The 1.0-point differences on Terminal-Bench and GPQA are noise; the 12-point gap on SWE-bench Pro is real, and Qwen acknowledges it internally.
The generational jump from Qwen3.7-Max is the more useful number. DeepSWE goes from 21.6 to 56.6, FrontierSWE from 40.7 to 73.5, and JobBench from 31.3 to 53.4. The internal Alibaba score index rose from 0.474 to 0.725 across the RL training run and topped out at roughly 4,000 environments before declining - a sign that the team is hitting the wall on synthetic environment scaling, not on model capacity.
What actually runs on your desk
The Max itself is not a desktop model. With 2.4 trillion total parameters and an expected 594 GB-class download at MXFP4 (the same envelope as Moonshot’s Kimi K3, per Apidog), it needs multi-node hardware and a serious KV-cache budget at 1M-token context. That is a hosted-API story, not a local one.
Qwen3.8-27B is the one that matters for self-hosting. Daniel Han, co-founder of Unsloth, posted on LinkedIn that “Qwen3.8-27B will run locally on 17GB RAM/VRAM setups and is expected to be the best performing model for its size,” and said Unsloth plans to support it. That claim has not yet been independently tested. Kimi K3’s 594 GB MXFP4 release, by comparison, is multi-node territory.
The 27B benchmarks have not been published yet, and Han’s “expected to be the best performing model for its size” is a forward-looking claim. The bar it has to clear is real: Apidog’s comparison describes Qwen3.6-27B as already widely seen as one of the strongest local models in its size class, and our own weekly open-weight showdown coverage showed the 27B dense checkpoint beating 397B-class MoE models on coding tasks in April.
What This Means
For people running models on a single GPU, Qwen3.8-27B is the more consequential of the two announced releases. If the 27B arrives next week and independent tests confirm Unsloth’s 17 GB claim, it could become a strong option for the 16-24 GB VRAM tier. The open question is how its unpublished benchmarks and real-world throughput compare with existing local models.
For the API market, Qwen3.8-Max is a price-pressure event. At $6 per million output tokens against Kimi K3’s $15, the same prompt and the same context cost less than half as much to run on Alibaba, and the 1M-token flat window removes the tiered-pricing trick that competitors use to bill long contexts. The Hong Kong market noticed immediately - Alibaba’s HK shares closed up 7% at HK$125.20 on the announcement day. The SCMP framing of “second only to Anthropic’s Claude Fable 5” matters because it comes from a Chinese outlet, not from Alibaba itself.
The licensing question is still open. GLM 5.2 went out under plain MIT, and the smaller Qwen releases have been Apache 2.0. Until Alibaba publishes the license for the 2.4T weights, enterprises that need self-hosting the flagship rather than the smaller sibling should wait for the August 10 drop before assuming what they can do with the model.
The Bottom Line
Qwen3.8-Max is one of the largest open-weight releases announced by a major lab, and the 27B sibling is the one aimed at single-machine users. If independent testing confirms Unsloth’s 17 GB claim after the weights arrive, the 16-24 GB local-AI tier will have a notable new option. Until then, Qwen3.8-Max is available as a hosted API and both open-weight releases remain announced rather than tested.