Alibaba Qwen3.8-Max Benchmarks: 2.4T Parameters Challenge Claude

Alibaba has fired its latest shot in the AI arms race with Qwen3.8-Max, a 2.4-trillion-parameter model it calls the largest and most capable release in the Qwen family to date, and the first Max-class model the company has promised to open-source1117. Unveiled on August 3, the model arrives with a claim that is becoming familiar from Chinese labs: it can stand toe-to-toe with the best American frontier models, at a fraction of the price1113.

Markets took the claim seriously. Alibaba's Hong Kong-listed shares rose as much as 7% on the announcement, with U.S. premarket trading adding another 4.5%11. But the more interesting story for developers and buyers is in the benchmark tables — where the picture is genuinely strong, genuinely mixed, and still awaiting independent verification.

The technical picture

Qwen3.8-Max is built on the Qwen 3.5 foundation and uses a sparse mixture-of-experts architecture with a hybrid attention mechanism1418. The headline numbers are striking: 2.4 trillion total parameters, but only 95 billion activated per query, a design that Alibaba says cuts latency and compute cost relative to dense models of comparable scale1118. The context window stretches to 1 million tokens, enough to process lengthy legal documents, entire code repositories, or long videos in a single prompt1319.

The model is natively multimodal, handling text, images and video, and Alibaba says it can process documents over 200 pages and videos longer than 100 hours1719. By total parameter count it is the second-largest open-weight model in the world, behind only Moonshot AI's Kimi K3 at 2.8 trillion — a notably crowded field in which Alibaba also owns roughly 36% of Moonshot1219.

Where the benchmarks land

Alibaba published an unusually broad benchmark table with the launch, comparing Qwen3.8-Max against Anthropic's Claude Fable 5, Claude Opus 4.8, and OpenAI's GPT-5.6 Sol56. The results split cleanly into two stories.

On agentic and productivity tasks, the model wins frequently. It beats GPT-5.6 Sol on five of seven agent benchmarks, including CoWorkBench (74.8 vs. 71.5), WorkSpaceBench (67.7 vs. 65.6) and JobBench (53.4 vs. 45.4)12. Its 93.0 on PaperBench — a research-reproduction benchmark — is the highest score in the comparison, ahead of GPT-5.6 Sol at 90.5 and Fable 5 at 88.856. On IFBench, an instruction-following test, its 82.8 dwarfs Fable 5's 63.5 and Opus 4.8's 62.26. And on OSWorld-Verified, which measures a model's ability to operate real desktop applications, Alibaba reports 86.1, ahead of Fable 5 at 85.0 and GPT-5.6 Sol at 83.258.

On the hardest reasoning and frontier-coding work, the lead changes hands. SWE-bench Pro comes in at 67.7 against Fable 5's 80.0 — a 12-point gap — and FrontierSWE shows an even wider spread, 73.5 versus 88.868. On Humanity's Last Exam, the broad multidisciplinary knowledge test, Qwen3.8-Max manages 43.6 while Fable 5 sits at 53.3 and GPT-5.6 Sol at 47.268. GPQA Diamond, graduate-level science reasoning, is essentially a tie: 92.6 for the Qwen model against 92.6 for Fable 5 and 94.1 for GPT-5.6 Sol56.

The pattern across coverage is consistent: Qwen3.8-Max wins broadly on agent, multimodal and document-intelligence benchmarks, but where deep single-shot reasoning or the very hardest coding challenges set the bar, Fable 5 still leads by 10 to 15 points1217. One analyst's blunt summary captures the consensus: it beats Opus 4.8 on several agentic rows, trails Fable 5 on most core coding work, and wins big on multimodal — a strong result, but not the across-the-board frontier victory a skim of the launch coverage might suggest6.

The long-horizon autonomy claims

The most distinctive part of the launch is Alibaba's emphasis on what it calls long-horizon tasks — sustained, self-directed work over days rather than single prompts1517. Alibaba reports a 16-day run in which the model autonomously built a self-evolving agent framework from scratch, racking up 265 commits, 127 pull requests and 151 issues with no human touch1718. In a second demonstration, it gave the model a research paper with no starter code; over roughly 125 hours, the model wrote 7,600 lines of code, ran 33 GPU training jobs, reproduced all six of the paper's main results, and then beat the original method on the AIME24 math benchmark by 2.7 points1217. In a simulated e-commerce business exercise seeded with 152 scammers to detect, the model quadrupled its starting capital to 416,252 yuan — 38% more than runner-up GLM 5.217.

These are vendor-reported demonstrations, and the usual discount applies. But the capability they describe — multi-day autonomous execution — maps directly onto the agent benchmarks where the model scores best, and it is the clearest differentiator from earlier Qwen Max tiers1217.

Independent signals and caveats

Early independent data is promising but not definitive. On the crowdsourced Arena.AI platform, Qwen3.8-Max immediately became the highest-ranked Chinese text model, though it still trails several Anthropic offerings; it ranked second globally on vision tasks, behind only a Fable 5 variant111319. Another tracking put it fifth in Text Arena and fourth in Frontend Code Arena, where its score of 1,668 sits just 37 points behind Claude Opus 514.

There are real caveats. Every score in Alibaba's comparison table comes from Alibaba's own runs — competitor numbers included — and full independent verification from evaluators like Artificial Analysis or LMArena was still pending as of the coverage812. Alibaba has published no score on SWE-bench Verified, the most widely tracked coding benchmark, where leading models score 93–97%8. And the practical barrier is formidable: a 2.4-trillion-parameter model occupies well over a terabyte even with heavy compression, keeping self-hosting firmly in data-center territory11.

There is also a history of gaps between vendor claims and measured results. When independent evaluators tested the earlier Qwen 3 Max, their accuracy came in roughly 11.6% below Alibaba's reported AIME25 score7.

The strategic read

The open-sourcing decision is the strategic headline. After keeping several recent flagship releases proprietary earlier in the year, Alibaba is returning to open weights at the top tier — a return timed precisely to the moment Moonshot, DeepSeek, ByteDance, MiniMax and Z.ai are all shipping increasingly powerful models1113. The pricing angle sharpens the challenge: at roughly $2 input and $6 output per million tokens, Qwen3.8-Max undercuts Fable 5 by a wide margin, with output-token pricing around a quarter of Claude Opus 5's61214.

My reading: this is not a frontier takeover — Fable 5 retains a clear lead on the hardest reasoning — but it is a genuine inflection in the open-weight race. Alibaba has produced the strongest open model on agentic and multimodal work, priced it aggressively, and staked its claim on the category many consider the next battleground: AI that works unsupervised for days at a time. The gap with the U.S. frontier is now measured in points, not years. Until independent verification lands, that is exactly what a serious contender looks like.