Published: August 4, 2026
Alibaba’s Qwen team released Qwen3.8-Max on August 3, 2026, making its new flagship model generally available through QwenCloud and Alibaba Cloud Model Studio ahead of an open-weights release next week. Qwen3.8-Max is a 2.4 trillion parameter Mixture-of-Experts model that activates 95 billion parameters per query, accepts text, image and video input, and supports a 1 million token context window at 2.00 dollars per million input tokens and 6.00 dollars per million output tokens. Alibaba is also open-weighting a second checkpoint, Qwen3.8-27B, which makes Qwen3.8-Max the first model in the Qwen-Max class to have its weights published.
What did Alibaba release with Qwen3.8-Max?
Qwen3.8-Max is Alibaba’s next-generation flagship language model and the most capable model in the Qwen family to date. It is a Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion parameters active per query, built on the Qwen3.5 architecture. It is natively multimodal: it accepts text, images and video as input and returns text. Alibaba first showed the model as a preview in mid-July 2026 and moved it to general availability on August 3, 2026, alongside a full benchmark table that had been missing from the preview.
The release covers two checkpoints. Qwen3.8-Max is the flagship, served today through the hosted API. Qwen3.8-27B is a smaller checkpoint aimed at ordinary on-premise GPU hardware. Weights for both are scheduled to go live on Hugging Face and ModelScope next week. That marks Alibaba’s return to open-sourcing its top-tier models after keeping several 2026 flagship releases proprietary, and it puts a Max-class Qwen model in open-weight territory for the first time. Alibaba shares rose roughly 7 percent in Hong Kong and 4.5 percent in New York premarket trading on the announcement.
Alibaba positions the model around long-horizon agentic work rather than single-turn answering. The stated target applications are repository-scale coding agents, long-document knowledge bases, long-video indexing, structured data extraction and multi-step research assistants. The company says the model can rebuild software applications from screenshots, generate interactive games and educational animations, and convert two-dimensional floor plans into 3D visualisations.
Qwen3.8-Max benchmarks and technical specs
Qwen3.8-Max lists a 1 million token context window.
Maximum input is 991,000 tokens, dropping to 983,000 when thinking is enabled.
Maximum output is 131,000 tokens in both modes and the maximum reasoning budget is 262,000 tokens.
Rate limits are 2 million tokens per minute and 15,000 requests per minute. A reasoning_effort parameter offers three levels that trade speed against thoroughness. Supported capabilities include function calling, structured outputs, batch requests, prefix completion and fine-tuning, and five built-in tools ship on the Responses API: code_interpreter, web_search, web_extractor, t2i_search and i2i_search.
On the benchmark table Alibaba published with the release:
- Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Claude Fable 5 at 84.6 and behind GPT-5.6 Sol at 88.8.
- It reports 67.7 on SWE-bench Pro against Claude Fable 5’s 80.0, and 73.5 on FrontierSWE against Claude Fable 5’s 88.8.
- It leads the comparison on PaperBench at 93.0 and IFBench at 82.8.
- GPQA Diamond lands at 92.6, up marginally from Qwen3.7-Max’s 92.4.
- The multimodal rows are stronger: OSWorld-Verified 86.1, Parametric CAD Bench 91.5 and OmniDocBench 1.5 at 92.1.
- Against its own predecessor the agentic jump is large, with DeepSWE 1.1 moving from 21.6 to 56.6, FrontierSWE from 40.7 to 73.5 and JobBench from 31.3 to 53.4.
Two caveats belong with those numbers. All figures are vendor-reported from internal runs and independent verification is still pending. And the multimodal table benchmarks Qwen3.8-Max against Qwen3.7-Plus rather than Qwen3.7-Max, which flatters the generational delta. Alibaba’s own reinforcement learning scaling curve is also worth reading closely: the internal score index across more than ten benchmarks rose from 0.474 to 0.725, but it peaks at around 4,000 training environments and then declines to 0.719 and 0.689.
What can Qwen3.8-Max do on long-horizon agent tasks?
Alibaba published five case studies in which Qwen3.8-Max worked without human intervention. In the first, the model spent 16 days building a command-line tool called oh-my-cli, turning incoming user requests into GitHub issues, assigning them to itself, writing the code and running tests. By July 30, 2026 it had produced 265 commits, 127 pull requests and 151 issues with no human commits. In the second, it was given the research paper “Unified Data Selection for LLM Reasoning” with no starter code, and over roughly five days and 125 hours of compute it wrote 7,600 lines of code, ran 33 GPU training jobs, reproduced all six main results and then beat the paper’s method on the AIME24 benchmark by 2.7 points.
The third case study entered the model into the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba’s Tianchi platform against 526 human teams. Within 24 hours and across 45 submissions, accuracy climbed from 0.60 to 0.853, placing it ahead of 458 of the 526 teams. The fourth had the model design a cryptographic circuit, reducing a working design from 8,298 logic gates to 678 over roughly 500 iterations and shrinking the physical area from 106 by 106 to 46 by 46 micrometres, an 81 percent reduction. The fifth, E-Commerce-Bench, simulates a full fiscal year of online retail on anonymised Taobao and Tmall data, including 152 hidden scam suppliers. Starting from 100,000 yuan, Qwen3.8-Max finished with 416,252 yuan, 38 percent ahead of runner-up GLM 5.2 and more than 2.5 times what Qwen3.7-Max managed.
On the multimodal side, Alibaba says the model handles documents over 200 pages and videos longer than 100 hours. The company is also introducing RecreationBench, a benchmark that requires a model to rebuild a running application without access to source code, observing the target only through clicks and keyboard input across Ubuntu, macOS, Windows, Android and the web. Alongside it, Alibaba is releasing Qwen-MM-Plugins, an extension library that adds image and video processing, visual tool use and multimodal memory to existing agent systems.
How does Qwen3.8-Max compare to Claude Fable 5, GPT-5.6 Sol and Kimi K3?
On Alibaba’s own tables, Qwen3.8-Max sits near but generally below the Western frontier on pure software engineering. GPT-5.6 Sol still leads Terminal-Bench 2.1 at 88.8 against 86.6, and Claude Fable 5 leads SWE-bench Pro at 80.0 against 67.7 and FrontierSWE at 88.8 against 73.5. Where Qwen3.8-Max leads is agentic breadth and multimodality: PaperBench at 93.0, IFBench at 82.8 and most of the vision rows. On the crowdsourced Arena.AI platform, Qwen3.8-Max became the highest-ranking Chinese model for text tasks on launch while still trailing several Anthropic models, including Claude Fable 5.
The nearer competitor is domestic. Moonshot AI’s Kimi K3 is a 2.8 trillion parameter multimodal Mixture-of-Experts model with a 1 million token context window, released with open weights in July 2026. Qwen3.8-Max is smaller in total parameters but arrives with a hosted API, a published benchmark table and a promised open-weights drop, which is a broader launch surface than Kimi K3 had. The practical difference for buyers is that Qwen3.8-Max supports both OpenAI’s Chat Completions format and Anthropic’s API protocol, so it drops into Claude Code, Codex, Qoder CLI, Qwen Code and OpenClaw without harness work. Alibaba says performance is roughly consistent across those harnesses rather than tuned for its own QwenWork environment.
Qwen3.8-Max availability, pricing and open weights
Qwen3.8-Max is available now through QwenCloud and Alibaba Cloud’s Model Studio APIs, through QwenWork, Alibaba’s workplace agent platform, and in Qwen Studio at chat.qwen.ai. The API is OpenAI- and DashScope-compatible, so integration is a base URL and model ID change. Pricing is 2.00 dollars per million input tokens and 6.00 dollars per million output tokens. Implicit cache reads cost 0.25 dollars per million tokens, explicit cache creation 2.50 dollars and explicit cache reads 0.17 dollars per million tokens. Cached input is eight times cheaper than fresh input, so prompt prefix stability affects cost more than prompt length does.
Self-hosting is a different question. At 2.4 trillion total parameters, every expert stays resident in memory even though only 95 billion activate per query, which makes the flagship checkpoint a multi-node datacentre artefact rather than something that fits a single server. Qwen3.8-27B is the checkpoint intended for ordinary on-premise GPU hardware. Alibaba has not yet published the licence terms for either checkpoint, and the weights for both are due on Hugging Face and ModelScope next week, so teams planning an on-premise deployment should wait for the licence before committing.
Full technical details on Qwen 3.8 are on the Qwen blog.
