Qwen3.7-Max
Editorial notes
Anunciado 19 mayo 2026 en Apsara Summit (Hangzhou). Proprietary agent foundation model (closed-weight). Pricing $2.50/$7.50 per MTok (cache $0.25). Context 1M, output 65K, reasoning nativo (enable_thinking/preserve_thinking). API-only via Alibaba Cloud Model Studio, compatible OpenAI + Anthropic protocols. Solo texto. Lidera GPQA-Diamond (92.4 vs Opus 4.6 91.3) y HLE (41.4 vs Opus 4.6 40.0). Highlight: 35h autonomous kernel optimization (10.0x geometric mean speedup vs Triton ref en hardware T-Head ZW-M890 no visto en training). Otros scores oficiales del blog: HMMT-2026-Feb 97.1, IMOAnswerBench 90.0, Apex 44.5, IFBench 79.1, MRCR-v2 128k 90.4, WMT24++ 85.8, MAXIFE 89.2, PolyMATH 86.5, MMLU-Redux 95.0, SuperGPQA 73.6, MCP-Atlas 76.4, Kernel Bench L3 1.98x/96%.
Spec sheet
- Empresa
- Alibaba
- Pais
- CN
- Tipo
- reasoning
- Release
- 2026-05
- Context
- 1.0M tokens
- Licencia
- proprietary
- Pricing (alibaba)
- $2.5/$7.5/M
- Slug
- qwen3-7-max
Benchmarks (8)
Reasoning 3
Coding 4
- 91.6LiveCodeBenchProblemas de coding contests en vivo de LeetCode/Codeforces.
- 80.4SWE-bench-VerifiedIssues reales de GitHub de 12 repos populares de Python.
- 69.7Terminal-Bench-2Terminal Bench v2 - tareas agenticas en CLI.
- 60.6SWE-bench-ProVersion profesional de SWE-bench con issues mas complejos.
Cite this model
BibTeX · APA
BibTeX
@misc{frontier-qwen3-7-max,
title = {Qwen3.7-Max},
author = {{Alibaba}},
year = {2026},
note = {Frontier Benchmarks AI atlas. Accessed 2026-06-10},
url = {https://frontierbenchmarks.com/models/qwen3-7-max}
} APA
Alibaba (2026). Qwen3.7-Max [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-10, from https://frontierbenchmarks.com/models/qwen3-7-max
Citation refleja la pagina del atlas, no el paper original del modelo. Para el paper, ve a la seccion "Recursos" arriba.