The September 2, 2026 snapshot of Qwen3.8-Max, Alibaba's hosted flagship. Qwen3.8-Max is the API version of the open Qwen3.8-2.4T-A95B weights, a 2.4T MoE with 95B active, and adds vision input, a non-thinking mode, a 1M-token default context and built-in tools. The first snapshot was announced on August 3, and this one updates it with deeper coding, multi-tool agent orchestration, and chart and document vision. Pricing is unchanged at $2 / $6 per million input/output tokens, and the qwen3.8-max endpoint has routed to it since September 5. Alibaba has not released weights for this snapshot. At its 2026 Apsara Conference Alibaba said a month of fully automated recursive-self-improvement runs (33 iterative cycles) lifted Qwen3.8-Max's Artificial Analysis score from 40 to 45 (v4.3.2).

Artificial Analysis scores it 34 on Intelligence Index v5.0, Alibaba's highest v5 reading until Qwen4-Max (43). AA lists it as proprietary. It sits above the best open-weights models (DeepSeek-V4.1-Flash-1017 32, GLM-5.3 32) and far behind Moonshot's Kimi K3.1 Pro (46), the top Chinese model. It is strongest on AA-Briefcase (Elo 1608, level with Gemini 4 Carbon) and AutomationBench-AA (56.2%), and weak on the science evals: 11.9% on Terminal-Bench-Science and 28.5% on the CritPt v2 hard subset. It is expensive to run for its score, at $10.02 per index task. AA scored it 45 on v4.3.2, and the August 3 snapshot (40 on v4.3.2) has not yet been rescored on v5.0.

Model Details

Architecture MOE
Parameters 2.4T
Active params 95B
Context window 1,000,000
AA Intelligence 34 was 45 on v4.3.2
License Proprietary (API)
Base model qwen3.8
frontiermoereasoningagenticcodingmultimodalproprietary

Related