Alibaba released Qwen3.8-Max to general availability on August 3, 2026, shipping a 2.4 trillion-parameter Mixture-of-Experts model with a 1-million-token context window, native text, image, and video input, and a published benchmark table that puts it ahead of Claude Fable 5, GPT-5.6, and Claude Opus 4.8 on several agentic and long-document tasks. The model is live now on QwenCloud at $2.00 per million input tokens and $6.00 per million output tokens, with OpenAI- and Anthropic-compatible endpoints available from day one. The one piece Alibaba hasn’t delivered is the open-weight release it promised for the week of August 10.<\/p>\n
Qwen3.8-Max is Alibaba’s new flagship model from the Qwen team, built on the architecture introduced with Qwen3.7-Max but scaled up and extended to multimodal input for the first time above the trillion-parameter mark. It is a sparse Mixture-of-Experts model with 2.4 trillion total parameters and roughly 95 billion active parameters per token, depending on which third-party reporting you read. The model supports a 1-million-token context window, broken down as 991K maximum input, 131K maximum output, and reasoning chains up to 262K. It accepts text, image, and video input and returns text output.<\/p>\n
The general availability release on August 3 closed three gaps that defined the preview three weeks earlier:<\/p>\n
It also added a day-one Anthropic Messages-compatible endpoint, so Claude Code and similar agent harnesses can switch to Qwen3.8-Max by repointing the ANTHROPIC_BASE_URL and setting ANTHROPIC_MODEL to qwen3.8-max, with no other workflow changes.<\/p>\n
Alibaba’s published table gives Qwen3.8-Max a mixed but competitive record against the current frontier:<\/p>\n
The pattern is consistent. Qwen3.8-Max wins on agentic computer-use and long-document tasks (OSWorld-Verified, PaperBench, IFBench) and holds its own on general terminal agentic work, but loses to Claude Fable 5 by 12 points on SWE-bench Pro, the harder professional coding benchmark. On broad-knowledge reasoning (HLE), it finishes last of the four models tested. Alibaba’s own Arena.ai citation puts Qwen3.8-Max at rank 5 on Text Arena, between four different Claude configurations.<\/p>\n
Every number in the benchmark table comes from Alibaba’s vendor-run evaluation. No independent evaluator such as Artificial Analysis or a third-party LMArena listing has copied the full table as of this writing.<\/p>\n
Alibaba is positioning Qwen3.8-Max less as a chatbot upgrade and more as an autonomous-agent flagship. The headline demonstration is oh-my-cli, a command-line agent framework the model built on its own: turning incoming requests into GitHub issues, claiming them through a state machine, writing code, running end-to-end tests, and merging its own pull requests. As of July 30 the run had produced 265 commits, 127 pull requests, and 151 issues over 16 days without human intervention. A live check of the repository showed 797 commits, 61 open issues, an Apache-2.0 license, and a commit merged shortly before publication, a more convincing agentic-coding demonstration than a static benchmark score.<\/p>\n
In research and competition tasks, Alibaba reports Qwen3.8-Max reproducing a published paper on data selection for LLM reasoning, writing roughly 7,600 lines of code and running 33 GPU training rounds over five days to land a +2.71 point improvement on AIME24 over the original paper’s method. In a separate 24-hour coding competition, the model’s entry reportedly beat 458 of 526 human teams, finishing in the 87th percentile.<\/p>\n
The model also processes documents past 200 pages and video past 100 hours using video memory graphs, and pairs GUI screen operation with visual feedback loops for verifying its own output, evaluated internally against Alibaba’s RecreationBench.<\/p>\n
Qwen3.8-Max is priced at $2.00 per million input tokens and $6.00 per million output tokens on QwenCloud, with cached input as low as $0.17 per million tokens on explicit reads. That undercuts Kimi K3’s $3.00\/$15.00 pricing by a wide margin and roughly matches the prior Qwen3.7-Max rate card at $2.50\/$7.50, despite the jump in scale. Rate limits are 2 million tokens per minute and 15,000 requests per minute.<\/p>\n
Access runs through QwenCloud (model ID qwen3.8-max), Alibaba Cloud Model Studio’s international scope, and the Vercel AI Gateway at zero markup (alibaba\/qwen3.8-max). A day-one Anthropic Messages-compatible endpoint lets Claude Code workflows repoint ANTHROPIC_BASE_URL to Alibaba’s DashScope endpoint and set ANTHROPIC_MODEL to qwen3.8-max with no other changes, the cheapest way to run a side-by-side comparison against Claude models already in an existing agent harness.<\/p>\n
This is the unresolved piece. Alibaba said weights for Qwen3.8-Max and a smaller Qwen3.8-27B would land on Hugging Face and ModelScope during the week of August 10. As of August 11, one day into that window, no repository has appeared for either model, and no license has been named. It remains an open question whether a Max-class open weight would ship under the permissive Apache-2.0 license used for smaller Qwen releases like Qwen3.6-27B, or something more restrictive. Until weights land, open_source: false is the accurate label for this model regardless of what has been promised.<\/p>\n
Not yet. Alibaba promised open weights for Qwen3.8-Max and a smaller Qwen3.8-27B during the week of August 10, 2026, but as of August 11 neither model has appeared on Hugging Face or ModelScope, and no license has been confirmed.<\/p>\n
$2.00 per million input tokens and $6.00 per million output tokens on QwenCloud, with cached reads as low as $0.17 per million tokens. That is roughly one-third the per-token cost of Kimi K3.<\/p>\n
It depends on the task. Alibaba’s own table shows it ahead on OSWorld-Verified, PaperBench, and IFBench, but behind Claude Fable 5 by 12 points on SWE-bench Pro and behind all three rivals on HLE. It is not a clean sweep in either direction.<\/p>\n