Qwen3.7 Max
Released May 19, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- China
- Type
- LLM
- Modality
- Text
- Context window
- 1M
- Max output
- 66K
- Speed
- 199 tok/s · 2.25s to first answer token
- Price ($/MTok in / out)
- $2.50 / $7.50
- Cost per task
- $1.03
- Open weights
- No
Benchmarks6 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
Humanity's Last Exam
AIME (latest)retired
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Long-horizon autonomy — runs for tens of hours and thousands of tool calls without hand-holding
- Precise instruction following, including multilingual and formatting constraints
- Heavy caching discounts make repetitive agent loops cheap in practice
Weaknesses
- Very verbose — a real agentic session costs far more than the per-token rate suggests
- Prose is over-engineered and reads machine-written
- Closed weights, a break from the rest of the Qwen line
SWE-bench Pro 60.6; Terminal-Bench 2.0 69.7. SWE-bench Verified 80.4 and HLE 41.4 lab/W&B-reported; MMLU-Pro 89.6 is third-party (LLM-Stats). Alibaba reports HMMT 97.1 instead of AIME. Terminal-Bench 2.1 74.5 is Alibaba's own figure from its Qwen3.8-Max comparison table (harness not stated; also on Together AI's Qwen3.7-Max page); vals.ai's independent harness scores it 61.05. Speed per Artificial Analysis' Qwen3.7 Max page (Alibaba's API), which carries no effort label.
For developers
API model strings
- Alibaba Cloud Model Studio
qwen3.7-max
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor