SWE-bench Pro across open-weight models hasn't moved in 10 weeks
Mar 27th - Jun 1st. 5 major open-weight releases. Net progress on SWE-bench Pro: +0.3%. Scores swung -5% and back within that window
Looks like open-weight labs are optimizing hard for agent tasks and multimodality – but not for coding anymore

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities
— MiniMax (official) (@MiniMax_AI) June 1, 2026
- Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
- MiniMax Sparse Attention scales context to 1M
-… pic.twitter.com/TF891iJukF
What SWE-bench Pro measures
SWE-bench Pro is Scale AI's coding benchmark for AI agents. Each task is a real issue from a professional code repository, and the model has to write a patch that passes the project's tests. The tasks are longer and harder than in the older SWE-bench Verified, so scores sit much lower.
Ten weeks without a new high
GLM 5.1 opened the window on March 27 with 58.4%. Three April releases followed. Kimi K2.6 scored 58.6%, shown as 59 in the table. MiMo V2.5 Pro scored 57.2%, and DeepSeek V4 Pro 55.4%. MiniMax M3 closed the window on June 1 with 59.0%.
The best open-weight score moved by less than one point in ten weeks. GPT-5.5 sat at the same level, 58.6%. Claude Opus 4.8 was far ahead at 69.2%.
The launch posts put the weight elsewhere. MiniMax led its M3 announcement with three things at once: coding, agent benchmarks such as Terminal Bench 2.1 and MCP Atlas, and a 1M-token context.
Update, October 19, 2026: the wall broke, then the benchmark changed
Two weeks after this post, Z.ai released GLM-5.2 with 62.1% on SWE-bench Pro, 3.1 points above MiniMax M3. It was the last major open-weight release to lead with that benchmark.
Kimi K3, DeepSeek V4 Pro 0813 and GLM-5.3 publish no SWE-bench Pro score in their model cards. They report DeepSWE, Terminal Bench and SWE-Marathon instead, which test longer, more agent-like coding work.
On DeepSWE, open-weight models now sit close to the closed frontier. The scores below come from the GLM-5.3 model card:
| Model | Weights | DeepSWE v1.1 |
|---|---|---|
| GPT-5.6 Sol | Closed | 72.7 |
| Claude Fable 5 | Closed | 69.7 |
| Kimi K3 | Open | 67.5 |
| GLM-5.3 | Open | 66.9 |
| DeepSeek V4 Pro 0813 | Open | 62.7 |
| Claude Opus 4.8 | Closed | 58.0 |
| GLM-5.2 | Open | 46.2 |
GLM-5.3 went from 46.2 to 66.9 in one release. Open-weight coding did not stop improving after June. The progress moved to benchmarks where it is still visible.
Sources
- GLM-5.1 model card – Z.ai, SWE-bench Pro 58.4%
- Kimi K2.6 model card – Moonshot AI, SWE-bench Pro 58.6%
- GLM-5.2 model card – Z.ai, SWE-bench Pro 62.1% and the June comparison table
- GLM-5.3 model card – Z.ai, DeepSWE v1.1 comparison
- Kimi K3 model card – Moonshot AI, DeepSWE and Terminal Bench results
- DeepSeek V4 Pro 0813 model card – DeepSeek, DeepSWE 62.7