Qwen3.8-Max

Shattered.io's September 7, 2026 article explicitly names both "Qwen3.8-Max" and its open-weight packaging "Qwen3.8-2.4T-A95B," directly matching the subject's model variant. The article documents that Qwen3.8-Max first went live through Alibaba's QwenCloud API on August 2, 2026, with the open-weight variant published The article situates the launch within a crowded September 2026 frontier-AI window in which OpenAI, Anthropic, and Google also shipped new models within days of each other, highlighting Alibaba's choice to release full weights under Apache 2.0 rather than gate everything behind an API. It notes that Alibaba's stated pr

Releasebot's Qwen release-notes index confirms Qwen3.8-Max as a 2.4-trillion-parameter MoE model (95B active) focused on coding, work, research, and long-horizon tasks, with multimodal support and open weights announced to arrive shortly after the release. The entry anchors the Qwen3.8-Max SKU name and distinguishes it For developers, the index surfaces the headline capability claims — comprehensive coding/work improvements, end-to-end complex task delivery, and a forthcoming open-weights drop — though it is a third-party aggregator rather than a primary Qwen post, so benchmark and infrastructure specifics should be verified against

Emergent's Qwen 3.8 benchmark breakdown reports Qwen3.8-Max's published wins (PaperBench 93.0, IFBench 82.8, Terminal Bench 2.1 86.6, MathVision 95.2, OSWorld-Verified 86.1) alongside losses (HLE 43.6 last among flagships, SWE-bench Pro 67.7, 12 points behind Fable 5), with explicit caveats that every figure is vendor- It also documents published pricing at roughly $2 input and $6 output per million tokens — about a third of Claude Opus 5 and a quarter of GPT-5.6 Sol — and flags that no independent evaluator (Artificial Analysis, community leaderboards) had scored the model as of August 14, 2026, making it essential reading for any d

Alibaba officially announced Qwen3.8-Max on August 3, 2026 as the most powerful model in its Qwen series to date, directly relevant to the subject's Qwen3.8-2.4T-A95B packaging. The model features 2.4 trillion total parameters with a Sparse Mixture-of-Experts architecture combined with a hybrid attention mechanism, act The release emphasizes strong performance in coding, real-life work, research, and long-horizon tasks, where Qwen3.8-Max is ranked fourth in Fronted Code Arena and demonstrates proficiency in autonomous coding and extended independent execution. The MoE-plus-hybrid-attention design is described as balancing massive sca

The alphaXiv mirror of Qwen's official Qwen3.8-Max blog describes a 2.4-trillion-parameter MoE model (95B active) marketed as the first Qwen-Max-class checkpoint with open weights (scheduled for release the following week). It positions frontier capability as multi-day autonomous execution across software engineering, For developers evaluating the Modal-hosted endpoint, the post signals strong coding-agent potential and a 1M-class context target, but the competitive comparisons and multi-day autonomous claims are vendor-authored and should be treated as marketing until independently verified on Modal's inference route.

EvoLink's post on Qwen3.8-Max confirms the upstream QwenCloud catalog entry "qwen3.8-max" as the model flagship and notes the August 3, 2026 addition to Qwen's release log. The post documents a 1M-token context window, native vision-language support, a 2.4-trillion-parameter Mixture-of-Experts architecture, and hybrid The article clarifies naming by stating Qwen3.8 Max is not Qwen3-8B and that Qwen3.8 is a 2026 model generation distinct from the older eight-billion-parameter Qwen3-8B checkpoint. As of the EvoLink write-up, open weights and a license were not yet verified as released, with the post noting only that Qwen announced a f