Coding · Benchmark profile
A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior.
Data verified 21 Jul 2026 · Methodology 1.6.0
Benchmark score on ProgramBench: Can Language Models Rebuild Programs From Scratch?
Kimi K3 leads at 77.8%, followed by GLM-5.2 (63.7%) and Kimi K2.7 Code (53.6%).
Visual analysis
Switch between model placement, score distribution and descriptive provider averages. Every view uses the same sourced leaderboard.
3 ranked models · higher is better
| Model | Rank | Provider | Score |
|---|---|---|---|
| Kimi K3 | #1 | Moonshot AI | 77.8% |
| GLM-5.2 | #2 | Z.AI | 63.7% |
| Kimi K2.7 Code | #3 | Moonshot AI | 53.6% |
One best score per model · higher is better
| Rank | Model | Provider | License | Evidence use | Score |
|---|---|---|---|---|---|
| #1 | Kimi K3 kimi-k3 | Moonshot AI | closed | Reference only | 77.8% |
| #2 | GLM-5.2 glm-5-2 | Z.AI | open | Estimated reference | 63.7% |
| #3 | Kimi K2.7 Code kimi-k2.7-code | Moonshot AI | open | Estimated reference | 53.6% |
The top of this snapshot is led by Kimi K3 at 77.8%; third place is 24.2 points behind. The top-3 spread is 24.2 points.
About ProgramBench: Can Language Models Rebuild Programs From Scratch?
A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior.
Kimi K3 by Moonshot AI currently leads with 77.8%.
3 models in the LuminaBench cohort have a qualifying score on this benchmark.
No. This benchmark is display-only and does not enter the overall Lumina composite.
Related
