MDL-6052EST.2025 · IDX.651
Language modelIn productionretires 10/10/2026 — in 30 days

Qwen3-Max

Alibaba Cloud · China · 2025

Alibaba's first trillion-parameter model and the flagship of its paid API for a year — closed-weight, trained on 36 trillion tokens, and switched off on 10 October 2026.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Qwen3-Max was the model with which Alibaba crossed the trillion-parameter line. Announced in preview at the start of September 2025 and released in its official snapshot on 23 September, it followed the Qwen3-2507 series and became, in the company's own words, its largest and most capable model. The published figures are unusually specific for a closed model: over one trillion parameters in a mixture-of-experts arrangement built on the Qwen3 design with a global-batch load-balancing loss, pretrained on 36 trillion tokens. Alibaba described a pretraining run that never needed a rollback — no loss spikes, no changes to the data mix — and credited its own infrastructure: PAI-FlashMoE pipeline parallelism for a 30 percent gain in model-FLOPs utilisation over Qwen2.5-Max, and a ChunkFlow strategy that tripled throughput on long sequences and made training at a one-million-token context length practical. In the API the model was served with a 256,000-token window, with thinking mode, function calling, built-in tools and structured output. On benchmarks the instruct version scored 69.6 on SWE-bench Verified and 74.8 on Tau2-Bench, the latter ahead of Claude Opus 4 and DeepSeek V3.1, and its preview held a top-three place on the LMArena text leaderboard, above GPT-5-Chat. A reasoning variant, Qwen3-Max-Thinking, was shown reaching 100 points on AIME 25 and HMMT when given a code interpreter and parallel test-time compute. Unlike most of the Qwen line, the weights were never published — Qwen3-Max existed only as a rented API on Alibaba Cloud Model Studio and in Qwen Chat. That is why its retirement is final rather than nominal: on 10 October 2026 Alibaba switches off qwen3-max together with its snapshots qwen3-max-preview, qwen3-max-2025-09-23 and qwen3-max-2026-01-23, and points customers to qwen3.7-plus. After that date the model cannot be run by anyone, including its author. The successor Max-class models — Qwen3.7-Max and Qwen3.8-Max, the latter the first Max ever released for download — have their own profiles in this catalogue.

#MoE#frontier#closed weights#China#retiring
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review