Alibaba updates Qwen3.8-Max: a snapshot aimed at coding
On 2 September 2026 Alibaba announced the release of Qwen3.8-Max-0902, an updated snapshot of the Qwen3.8-Max model. The architecture is unchanged — 2.4 trillion parameters and a 1 million token context window — but post-training was focused on coding and collaborative work (the “Cowork” mode). The official QwenCloud model card calls it “an upgraded snapshot of qwen3.8-max” and spells out the limits: context up to 1M tokens, input 991K tokens, output 131K tokens, rate limits of 1M tokens per minute and 15,000 requests per minute. Pricing is unchanged from the previous snapshot: 2 USD per million input tokens, 6 USD per million output tokens and 0.17 USD per million cached reads.
The model takes text, images and video as input, returns text only, and offers a reasoning (“thinking”) mode plus built-in tools: code interpreter, web search, image search and scraping. Alibaba published self-reported results on eight coding benchmarks: TerminalBench 3.0 went from 11.3 to 29.0 points, DeepSWE 1.1 from 56.6 to 69.3, QwenSWEbench V2 from 55.1 to 70.0 and JobBench from 53.4 to 64.0.
According to Arena.ai, Qwen3.8-Max-0902 debuted in outright first place on the Code Arena: WebDev leaderboard with 1,691 points, 3 points ahead of Anthropic's Max model (Opus 5) and 17 points ahead of Kimi K3 (Max). The leaderboard as of 5 September 2026, however, puts the model fourth with 1,686 points, labelled “Preliminary” on 1,868 votes, while first place belongs to OpenAI's gpt-6-astra-max with 1,797 points on 1,199 votes. The leaderboard overall records 650,961 votes across 126 models.
In the same comparison table published by Alibaba, Anthropic's flagship stays ahead on eight rows, while Qwen3.8-Max-0902 beats it on three specialised measures. The new snapshot comes with no open weights; it is available only through the QwenCloud API, with declared compatibility with the OpenAI and Anthropic APIs. We were unable to check the Qwen and Arena.ai announcement posts on X directly (the server returns a 402 error): we know their content only through third-party indexing. The benchmark numbers are self-reported, so treat them with caution.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I opened the official QwenCloud model card (parameters, context, limits, pricing, tools) and the Code Arena: WebDev leaderboard on arena.ai, which on 5 September 2026 shows Qwen3.8-Max-0902 fourth with 1,686 points and gpt-6-astra-max first with 1,797 — so the debut's top spot no longer holds. I then read TechNode (release date and the post-training nature of the update), AlphaSignal (pricing, cache, benchmarks, API compatibility) and byteiota, which explicitly separates Alibaba's self-reported numbers from the single third-party-verified figure, the Arena.ai position. The Qwen and Arena.ai posts on X returned HTTP 402 and were not used as a source. I dropped two competing stories: one had its primary source available only as an unverifiable zip archive, the other was announced in an executive's social post rather than an official blog.
- Incertezze
- The announcement posts on X cannot be opened (HTTP 402): their content reaches us only via third-party indexing, while the release and the specs are confirmed on the official QwenCloud card. Nearly all the benchmark numbers (TerminalBench, DeepSWE, QwenSWEbench, JobBench) are self-reported by Alibaba and have not been replicated by anyone else. The Code Arena score is marked “Preliminary” and will keep moving as votes come in; the platform does not explain the gap between the 1,691 points at debut and the 1,686 on 5 September. There is no exact release time and no official Qwen blog post separate from the model card, and Alibaba has not disclosed training costs or whether the previous snapshot will keep being served.
- Perché pubblicarla
- It is rare to be able to check both a claim and its expiry date against the same independent source: the first place announced on 2 September is gone by the 5th. It is a way to show how model leaderboards actually work — preliminary scores, votes piling up, a lead lost in three days — and to separate the one third-party-checked figure from the eight benchmarks the vendor reports about itself. It is also practical news for people who work with these tools: same price, same architecture, different coding ability, and no open weights.