Model Drop

Qwen3-Max

September 24, 2025 2 min read

Qwen3-Max, which Alibaba released on September 24, 2025 after a preview on September 5, is its largest model to date, a closed-weights model with over a trillion parameters built for coding and agent work.

Model: Qwen3-Max (qwen3-max, snapshot qwen3-max-2025-09-23)
Maker: Alibaba
Ship date: September 24, 2025 (preview September 5, 2025)
Pricing: $1.20 / $6 per M tokens (international list price, inputs up to 32K; higher above that)
Weights: Closed
Where you can use it:
- Qwen Chat
- Alibaba Cloud Model Studio API
When to reach for it:
- Agentic coding and tool calling
- Fast answers with no thinking pass
- Skip it for long step-by-step reasoning
How it compares:
- Beats Claude Opus 4 on tool use, per Alibaba
- Trails Claude Opus 4 on real-repo bug fixes
What makes it different:
- Over 1T parameters, trained on 36T tokens
- Alibaba's biggest model, weights kept closed
Early reactions: Reviewers saw a capable closed flagship that, by Alibaba's own tests, generally trailed Gemini 2.5 Pro and GPT-5.
Read more: Qwen3-Max: Just Scale it

The benchmarks

Qwen3-Max-Instruct, a non-thinking model; Alibaba compared it with its own open Qwen3-235B-A22B-Instruct-2507 and with Claude Opus 4 and DeepSeek V3.1 in non-thinking mode, not with thinking flagships like GPT-5 or Gemini 2.5 Pro, and the τ²-Bench figure is a weighted average across its domains. Alibaba also showed 100% on AIME 25 and HMMT for Qwen3-Max-Thinking, but that variant was still in training and wasn't released.

Alibaba's numbers

SWE-bench Verified (real GitHub fixes): 69.6%. +17.4 over Qwen3-235B-A22B-Instruct-2507 (52.2%), −2.9 under Claude Opus 4 (non-thinking) (72.5%), +3.6 over DeepSeek V3.1 (non-thinking) (66.0%).
τ²-Bench (Weighted) (agent tool calling): 74.8%. +21.9 over Qwen3-235B-A22B-Instruct-2507 (52.9%), +7.1 over Claude Opus 4 (non-thinking) (67.7%), +28.4 over DeepSeek V3.1 (non-thinking) (46.4%).
LiveCodeBench (v6, Feb to May 2025 problems): 69.0%. +17.2 over Qwen3-235B-A22B-Instruct-2507 (51.8%), +24.4 over Claude Opus 4 (non-thinking) (44.6%), +16.7 over DeepSeek V3.1 (non-thinking) (52.3%).
AIME 2025 (competition math): 81.6%. +11.3 over Qwen3-235B-A22B-Instruct-2507 (70.3%), +47.7 over Claude Opus 4 (non-thinking) (33.9%), +31.8 over DeepSeek V3.1 (non-thinking) (49.8%).
SuperGPQA (graduate-level knowledge): 65.1%. +2.5 over Qwen3-235B-A22B-Instruct-2507 (62.6%), +8.6 over Claude Opus 4 (non-thinking) (56.5%), +5.3 over DeepSeek V3.1 (non-thinking) (59.8%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 15.6. −3.0 under Claude Opus 4.1 (non-reasoning) (18.6), +1.9 over DeepSeek V3.1 (non-reasoning) (13.7), +0.3 over Kimi K2 0905 (15.3).