Mirr's coding model. It can answer a question, but it can also keep going: reading your repository and carrying the work through inside Mirr.
Mirr 1.0 Pro| Model ID | mirr-1.0-pro |
|---|---|
| Default | The default with no configuration. setup prepares it. |
| Best for | Repository investigation · debugging · multi-file changes · planning · review · long-running work |
| Languages | Keeps sessions that mix Korean requests with English symbols. |
| Operated by | Post-trained and operated by Patty. |
| Evaluation | DeepSWE +7.0pp over the pre-post-training checkpoint |
| Access | Runs inside Mirr, no separate account. |
Every model solved the same 200 Korean tasks four times, in the same Mirr workspace with the same tools, budget and grading.
| Overall (200) | Java bug fixes (75) | JavaScript bug fixes (75) | Python retention (50) | |
|---|---|---|---|---|
| Claude Opus 5◎ | 95.4 | 97.3 | 98.0 | 88.5 |
| Kimi K3 | 94.8 | 95.7 | 99.0 | 87.0 |
| Mirr 1.0 Pro | 92.4 | 92.3 | 97.0 | 85.5 |
| GLM-5.3 | 92.4 | 93.3 | 96.7 | 84.5 |
| GPT-5.6 Sol◎ | 88.5 | 87.3 | 85.7 | 94.5 |
| MiniMax-M3 | 86.2 | 84.0 | 95.3 | 76.0 |
| DeepSeek v4.1 Flash | 73.2 | 75.3 | 82.0 | 57.0 |
| tasks | Mirr 1.0 Pro | Claude Opus 5◎ | Kimi K3 | GLM-5.3 | GPT-5.6 Sol◎ | MiniMax-M3 | DeepSeek v4.1 Flash |
|---|---|---|---|---|---|---|---|
| Arrays, lists, grids (36) | 97 | 99 | 97 | 94 | 99 | 90 | 80 |
| Strings and text (35) | 92 | 98 | 99 | 97 | 86 | 89 | 84 |
| Numbers and counting (43) | 96 | 99 | 97 | 97 | 85 | 88 | 76 |
| Bits, encodings, ciphers (12) | 96 | 98 | 94 | 96 | 100 | 92 | 77 |
| Real-world formulas (49) | 85 | 88 | 89 | 86 | 81 | 79 | 64 |
| Units, geometry, time (25) | 93 | 94 | 95 | 86 | 92 | 86 | 61 |
Mirr 1.0 Pro was trained on Korean. DeepSWE: +7.0pp over the pre-post-training checkpoint.
SWE-bench, Terminal-Bench, and DeepSWE are useful views of coding agents. Their tasks and interactions, however, are mostly English. Patty re-runs the same tasks with Korean requests and publishes each score together with its conditions.
| Evaluation | What it measures | Disclosure | Status |
|---|---|---|---|
| DeepSWE | Real repository issue resolution | Pre-post-training checkpoint vs Mirr 1.0 Pro, matched conditions | Internal matched eval · +7.0pp |
| SWE-bench | GitHub issue patch resolution | Exact variant, subset, and harness disclosed | Korean comparison in review |
| Terminal-Bench | Multi-step work in a terminal | Version, environment, limits, and tools disclosed | Korean comparison in review |
| Patty Korean Sessions | Long mixed-language coding sessions | Follow-ups, context retention, and verification evaluated | Protocol in review |
A benchmark episode is not a repository you work in for days. Translation equivalence, runtime environment, and grader variance can all move a score. Patty publishes the conditions next to the number, or does not publish the number.
DeepSWE · lift over pre-post-training checkpoint
Yes. Patty post-trains and operates Mirr 1.0 Pro for software work and Korean collaboration.
No. Code is mixed-language by nature. Move between Korean and English as needed.
No. Ask it to investigate, explain, plan, or review without producing a patch.
Mirr 1.0 Pro is the model making judgments. Mirr is the workspace that connects it to files and tools.
Task language, harness, tools, time limits, and multi-turn context all change the result.
