Alibaba launches Qwen4-Max: 1M context MoE model tops Chinese models on RSI-Exam
HuaxiuYaoML · x · 2026-11-17
Alibaba's Qwen team released Qwen4-Max, built on a new MoE architecture with native vision and a 1M-token context window.
Official highlights:
- 50% fewer tokens per task vs Qwen3.8-Max, with per-task cost down over 60% in internal evals
- Stronger professional tool use, long-horizon execution, and deliverable quality across 3D modeling, game dev, manufacturing, finance, law, and healthcare
- Pricing: $1.5/M input tokens, $4.5/M output tokens, $0.075/M implicit caching
Benchmark: On RSI-Exam, which tests long-horizon recursive self-improvement over 88 tasks, Qwen4-Max scored 0.4425 (#6 overall), surpassing GPT-5.6-sol (0.4331) as the top Chinese model — up from #10 (0.3923) for Qwen3.8-Max-0902. The leaderboard is topped by Claude Opus 5.5 (0.5357), GPT-6-astra (0.5126), and Gemini 4 Argon (0.4943).
More from coding & agent
- YC amplifies viral take: 99.999% of 2023 'software development' is automated, yet SWE openings are up — ycombinator · 2026-11-17
- Agent orchestration: weak model at front desk, strong models in worker threads — dotey · 2026-11-17
- Huawei's EvoTrias: separating search, proposal, and judgment powers sets new SOTA — rohanpaul_ai · 2026-11-17
- Huawei's EvoTrias: separating search, code and scoring cuts agent token cost 64%, lifts OR-Tools 13.2% — rohanpaul_ai · 2026-11-17
- agent-browser: open-source browser automation so agents can verify fixes and file PRs — evilrabbit_ · 2026-11-17
- OpenRouter compares the context-token cost of MCP, CLI and Skills — rseroter · 2026-11-17