Alibaba says Qwen3.8-Max open weights ship next week

• Qwen3.8-Max left preview as Alibaba's new flagship: scaling01 described it as a 2.4T MoE with 95B active parameters, and Alibaba's official release says it supports up to 1M context.
• Open weights are scheduled for next week, and Qwen3.8-27B is also going open-weight, according to Alibaba_Qwen's launch thread.
• List pricing lands at $2/M input and $6/M output, with $0.25/M implicit cache in Alibaba_Qwen's pricing post; arena's Pareto chart put it on the Frontend Code Arena cost-performance frontier.
• The benchmark sheet is strong but footnote-heavy: cedric_chee warned that harness differences can move agentic scores, while arena marked the Text Arena result preliminary.
• Early hands-on reports are less absolute than the leaderboard: emollick's shader test called Qwen 3.8 Max solid but below Kimi K3 in his experiments.

Alibaba's official release says Qwen3.8-Max autonomously ran a 16-day software project and open-sourced the result as oh-my-cli. The QwenCloud model page quietly lists 991K max input, 131K max output, 262K max reasoning, 2M TPM, and 15K RPM. A Qwen Code pull request from the preview era treated qwen3.8-max-preview as mandatory-thinking because the endpoint rejected reasoning opt-outs.

• Model: qwen3.8-max, now out of preview and available through Qwen Studio and API access in Alibaba_Qwen's launch thread.
• Weights: Qwen3.8-Max weights are scheduled for next week, and Qwen3.8-27B is also going open-weight, according to Alibaba_Qwen.
• Shape: Alibaba's official release says the model uses Sparse MoE plus hybrid attention, with 2.4T total parameters and 95B active parameters.
• Context: the QwenCloud model page lists 1M context, 991K max input, 131K max output, and 262K max reasoning.
• Price: Alibaba_Qwen's pricing post lists $2/M input, $6/M output, and $0.25/M implicit cache; the QwenCloud page also lists $2.50/M explicit cache creation and $0.17/M explicit cache reads.
• License: Alibaba's official release gives the release window but does not name a license, which is why natolambert framed adoption as pending license and cedric_chee said he was hoping for MIT.

First-party

• TerminalBench-2.1: Qwen3.7-Max 74.5 → Qwen3.8-Max 86.6, +12.1 points, per Alibaba_Qwen's LM table.
• PaperBench: Qwen3.7-Max 64.8 → Qwen3.8-Max 93.0, +28.2 points, per Alibaba_Qwen's LM table.
• FrontierSWE: Qwen3.7-Max 40.7 → Qwen3.8-Max 73.5, +32.8 points, per Alibaba_Qwen's LM table.
• JobBench: Qwen3.7-Max 31.3 → Qwen3.8-Max 53.4, +22.1 points, per Alibaba_Qwen's LM table.
• Agents' Last Exam score: Qwen3.7-Max 31.1 → Qwen3.8-Max 52.4, +21.3 points, per Alibaba_Qwen's LM table.
• OSWorld-Verified: Qwen3.7-Max 73.3 → Qwen3.8-Max 86.1, +12.8 points, per Alibaba_Qwen's VL table.
• Vision2Web: Qwen3.7-Max 42.1 → Qwen3.8-Max 69.0, +26.9 points, per Alibaba_Qwen's VL table.

Third-party evaluators

• Frontend Code Arena: Qwen3.8-Max scored 1,668, -8 points versus Kimi K3 Max at 1,676 and -1 point versus Claude Opus 5 High at 1,669, according to arena's leaderboard thread.
• Text Arena: Qwen3.8-Max scored 1,496, -13 points versus Claude Fable 5 High at 1,509, according to arena's Text Arena result.
• Vision Arena: Qwen3.8-Max scored 1,305, -13 points versus Claude Fable 5 High at 1,318, according to arena's Vision Arena result.
• Frontend cost-performance: arena's Pareto chart placed Qwen3.8-Max on the frontier with Claude Opus 5, Kimi K3, GLM-5.2, and DeepSeek V4-Flash at $2/$6 pricing.

The first-party table also shows real gaps against other frontier models. SWE-bench Pro puts Qwen3.8-Max at 67.7 versus Fable 5 at 80.0, -12.3 points, while DeepSWE 1.1 puts it at 56.6 versus GPT 5.6 Sol at 73.0, -16.4 points, per Alibaba_Qwen's LM table.

Alibaba's official release says Qwen3.8-Max is built on Qwen 3.5, uses Sparse MoE plus hybrid attention, and activates 95B of its 2.4T parameters at inference time.

The QwenCloud model page lists image, text, and video input with text output. It also lists prefix completion, function calling, cache, structured outputs, batches, web search, and fine-tuning as supported features.

The preview-era Qwen Code pull request records one integration gotcha: qwen3.8-max-preview was treated as mandatory-thinking on DashScope-compatible routes, side queries stopped forcing enable_thinking: false, and structured requests avoided tool_choice: required because thinking mode rejected it.

The benchmark methodology caveat is unusually concrete. cedric_chee pointed to detailed footnotes and warned that different harnesses can materially affect agentic benchmark scores independent of model quality.

On a shader test, emollick called Qwen 3.8 Max solid but not Kimi K3 level in his experiments so far.

cedric_chee first put the preview's 3D and visual output close to K3 in his launch note, then said the GA model showed almost no visual-quality change from preview.

A math probe from suchenzang found a familiar split: formatting reminiscent of Sol, but reasoning laziness reminiscent of Opus.

Some agentic scores were closer to the new DeepSeek V4-Flash than the model size suggests, teortaxesTex argued, while still noting Qwen3.8-Max has vision and the scale of a frontier MoE.

• Alibaba: Alibaba_Qwen's launch thread points builders to Qwen Studio and API access, while Alibaba's official release says the model is accessible through Alibaba Cloud Model Studio and QwenWork.
• vLLM: vllm_project said it will support the Qwen3.8-Max open-source release on day 0.
• Local 27B: UnslothAI said Qwen3.8-27B should run locally on 17GB RAM or VRAM setups, after UnslothAI's earlier reply said the team was ready to make quants.
• Cline: Alibaba replied to Cline with “Can't wait to see Qwen3.8 in Cline soon” in Alibaba_Qwen's Cline reply.
• Hermes Agent: Teknium said Alibaba featured Hermes Agent in the release video.
• Qwen Code: the QwenLM pull request already patched preview support for DashScope side queries before the GA announcement.

• Qwen is opening a Max-class model for the first time, and cedric_chee called it the second open-weight model above 2T parameters after K3.
• Cost pressure is now part of the frontier story: cedric_chee estimated Qwen3.8-Max at nearly 3x cheaper than Kimi K3, 5 to 8x cheaper than Fable 5, and 2.5 to 4x cheaper than Opus 5.
• Adoption pressure is not only about scores. natolambert said Qwen's previous large API models did not get much adoption, and that pricing, license, release consistency, and patching feedback now separate Kimi, GLM, Qwen, and DeepSeek.
• The release landed in a crowded open-frontier week: simonw noted Qwen 3.8 Max and MiniMax-H3 arrived within hours, and natolambert called it another frontier open-weight model week.