Alibaba released Qwen3.8-Max to general availability on August 3, 2026, two weeks after previewing it at WAIC Shanghai with a single slide and a claim that it trailed only Claude Fable 5. The production launch shipped with what the preview lacked: a published benchmark table, per-token pricing, and a live production API. The missing piece is the open-weight release, which Alibaba said would land during the week of August 10 and, as of August 11, has not appeared on Hugging Face or ModelScope.<\/p>\n
Qwen3.8-Max is a 2.4 trillion-parameter Mixture-of-Experts model with roughly 95 billion active parameters per token, a 1-million-token context window (991K max input, 131K max output, with reasoning chains up to 262K), and native text, image, and video input with text output. It builds on the architecture introduced with Qwen3.7-Max and is the first Max-class model above the trillion-parameter mark to ship with multimodal input. The model ID on QwenCloud is qwen3.8-max.<\/p>\n
Alibaba published a four-way comparison table for the GA release. Every number is Alibaba’s own vendor-run evaluation; no independent evaluator has copied the full table yet.<\/p>\n
The pattern is consistent across the table: 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 Fable 5 by 12 points on SWE-bench Pro, the harder professional coding benchmark, and finishes last of the four on broad-knowledge reasoning (HLE).<\/p>\n
Long-horizon autonomous coding is the headline use case. Alibaba’s demo is oh-my-cli, a command-line agent framework Qwen3.8-Max built entirely on its own. The workflow turns incoming requests into GitHub issues, claims them through a state machine, writes code, runs end-to-end tests, and merges 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. The repository was still active when checked for this story, showing 797 commits, 61 open issues, an Apache-2.0 license, and a commit merged 33 minutes before publication.<\/p>\n
Research reproduction is another showcased capability. Alibaba reports Qwen3.8-Max reproduced 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 method. In a separate 24-hour coding competition, the model’s entry reportedly beat 458 of 526 human teams, landing in the 87th percentile.<\/p>\n
Native multimodal agents at scale round out the pitch. Qwen3.8-Max processes documents past 200 pages and video past 100 hours using what Alibaba calls 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 live now on QwenCloud at $2.00 per million input tokens and $6.00 per million output tokens, with cached input pricing from $0.25\/M on cache miss to $0.17\/M on explicit reads. That rate undercuts Moonshot’s Kimi K3 ($3.00 input, $15.00 output) by a wide margin and roughly matches Qwen3.7-Max’s prior $2.50\/$7.50 card, despite the jump in scale. Access is available through QwenCloud, Alibaba Cloud Model Studio’s international scope, and the Vercel AI Gateway at zero markup under alibaba\/qwen3.8-max.<\/p>\n
Alibaba shipped a day-one Anthropic Messages-compatible endpoint, so Claude Code can point at Qwen3.8-Max by changing ANTHROPIC_BASE_URL and ANTHROPIC_MODEL with no other workflow changes. That makes it the cheapest way to run a side-by-side comparison against Claude models inside an existing agent harness. Rate limits sit at 2M tokens per minute and 15,000 requests per minute.<\/p>\n
The open-weight release 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, no repository has appeared for either model, and no license has been named. Until weights land, the accurate label for this release is proprietary, not open-source. Smaller Qwen releases such as Qwen3.6-27B shipped under Apache-2.0, but whether a Max-class open weight would follow the same permissive license or something more restrictive remains an open question.<\/p>\n
Three other caveats apply. The 95-billion active parameter figure comes from third-party reporting, not from an Alibaba-published technical report or model card. Every benchmark in the table is Alibaba’s own vendor-run evaluation; independent evaluators like Artificial Analysis or LMArena have not yet copied the full table. And on HLE, Qwen3.8-Max finishes nearly 10 points behind Fable 5, so the launch is not a clean sweep in either direction.<\/p>\n
Strengths: a real, checkable benchmark table replaces the preview’s unverified marketing line; the model leads the four-way comparison on OSWorld-Verified, PaperBench, and IFBench; the $2\/$6 rate undercuts Kimi K3 while adding native multimodal input; the day-one Anthropic-compatible endpoint makes it a drop-in swap for Claude Code; and the oh-my-cli project is a live, ongoing, publicly auditable demonstration rather than a one-time benchmark run.<\/p>\n
Weaknesses: Qwen3.8-Max trails Claude Fable 5 by 12 points on SWE-bench Pro, finishes last of four flagships on HLE, ships only with vendor-run benchmark numbers, has missed its stated open-weight window without a confirmed license, and has not published an active-parameter figure in its own technical report.<\/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 of Kimi K3’s per-token cost.<\/p>\n
It depends on the task. Alibaba’s own table shows it ahead on OSWorld-Verified, PaperBench, and IFBench, but behind Fable 5 by 12 points on SWE-bench Pro and behind all three rivals on HLE.<\/p>\n