- input
- $0.80
- cached
- $0.40
- output
- $1.00
- Context
- 128Ktokens
- Released
- 1y ago
- Size
- 72Bparameters
Cost/1M tokens
Benchmarks
DeepSWE is the primary coding benchmark where an exact model evaluation exists. Arena leaderboards and other sourced evaluations appear below; missing scores are not zero.
How it ranks
Against its predecessor, similarly priced models and the models ranked around it.
DeepSWE 1.1
Best pass@1 per model · Higher is better
Not evaluated in DeepSWE 1.1
Arena Text
Best Arena rating per model · Higher is better
Not ranked on Arena Text yet
Arena WebDev
Best Arena rating per model · Higher is better
Not ranked on Arena WebDev yet
Arena Vision
Best Arena rating per model · Higher is better
Arena #131
- 1266
Gemini 3.5 Flash LiteSimilar price - 1234
MiMo-V2.5Similar price - 1160
Mistral Medium 3.1Similar price - 1153
GLM 4.5VSimilar price - 1127
Claude Haiku 3.5Nearby rank - 1127Mistral Small 3.1 24BNearby rank
- 1121
Qwen2.5 VL 72B Instruct - 1118
GPT-4o (2024-08-06)Nearby rank - 1113
Molmo 2 8BNearby rank
Ahead of 0 of 4 similarly priced models
Arena Search
Best Arena rating per model · Higher is better
Not ranked on Arena Search yet
Agent Arena
Best IPS score per model · Higher is better
Not ranked on Agent Arena yet
OSWorld
Best score per model · Higher is better
#10 of 10
- 63.3%
Kimi K2.5Similar price - 35.8%Claude Sonnet 3.7Nearby rank
- 23.0%o3Nearby rank
- 5.0%Qwen2.5 VL 72B Instruct
Ahead of 0 of 1 similarly priced model
Not evaluated in DeepSWE 1.1 as of 2026-09-22. Results are only shown for exact model versions in the official dataset. Official leaderboard.
All published scores
- Arena VisionArena #1311121
- default112195% CI 1110 to 1132 · 3,768 votes
- OSWorld#10 of 105.0%
- default5.0%qwen2.5-vl-72b-instruct (100 steps)
- default4.4%qwen2.5-vl-72b-instruct (15 steps)
Benchmark compilation: Epoch AI, Capabilities & benchmarking (CC BY attribution). Scores are converted to percentages and model names matched to the Vertical Tabs catalog. Individual evaluations are credited to their original sources. External results may carry separate licences and evaluation configurations; do not interpret scores from different harnesses as directly comparable.
Arena ratings: Arena (lmarena-ai/leaderboard-dataset), licensed CC BY 4.0. Ratings come from pairwise human votes and are only comparable within one leaderboard; the interval is Arena's 95% confidence interval. Model names are matched to the Vertical Tabs catalog.