Introduction
We are proud to introduce OxCoder-9B, our lightweight coding model for long-horizon tasks, agentic coding, and agentic reasoning. Despite featuring only 9 billion parameters, OxCoder-9B marks a substantial leap in agentic capabilities — particularly in Agentic Terminal and Agentic Coding tasks — punching far above its weight class and rivaling models many times its size. This release represents our strongest commitment yet to delivering frontier-level agentic performance in a compact, efficient, and openly accessible package.
- Trained on Frontier Agent Traces: Built from Fable-5.1 and GLM-5.3 agentic coding trajectories across Claude Code, OpenCode and Codex — OxCoder-9B has been distilled from some of the most capable agentic systems available, giving it a unique advantage in real-world software engineering scenarios.
- 262K Native Context: Equipped with a full 262,144 token native context window, enabling the model to handle complex, multi-file codebases and long-horizon reasoning tasks with ease.
- Error Recovery: Learns read-before-write patterns, responds to LSP diagnostics, and applies minimal edit diffs instead of full rewrites — making it robust in iterative coding environments where precision matters.
- Front-end Reasoning: OxCoder-9B exhibits remarkably strong front-end reasoning capabilities for its size, demonstrating a deep understanding of UI logic, component architecture, and web-native patterns that is rare in sub-10B models.
Benchmark
| OxCoder-9B | Ornith-1.5-9B | Ornith-1.0-9B | Qwen3.5-9B | Gemma-4-31B | |
|---|---|---|---|---|---|
| Coding | |||||
| Agentic terminal codingTerminal-Bench 2.1 (Terminus-2) | 49.6 | 46.2 | 43.1 | 21.3 | 42.1 |
| Agentic terminal codingTerminal-Bench 2.1 (Claude Code) | 50.8 | 47.0 | 40.6 | 18.9 | — |
| Agentic codingSWE-bench Verified | 73.5 | 70.6 | 69.4 | 53.2 | 52.0 |
| Real-world software engineeringSWE-bench Pro | 49.1 | 47.5 | 42.9 | 31.3 | 35.7 |
| Repo-level code generationNL2Repo | 36.2 | 32.4 | 27.2 | 16.2 | 15.5 |
| Reasoning | |||||
| Expert-level reasoningHLE (no tools) | 21.2 | 20.2 | 16.8 | 14.7 | 19.5 |
| Tool-augmented reasoningHLE (with tools) | 32.8 | 30.5 | 26.4 | 24.5 | 26.5 |
| Scientific reasoningGPQA Diamond | 86.9 | 86.4 | 82.5 | 81.7 | 84.3 |
| Agentic | |||||
| Multi-tool orchestrationMCP-Atlas | 56.7 | 54.2 | 49.4 | 46.8 | 55.0 |
| Web browsing and researchBrowseComp | 57.4 | 56.4 | 44.8 | 41.5 | — |
| Real-user agentic codingClawEval | 67.8 | 66.5 | 63.1 | 53.2 | 48.5 |
* All results reported for OxCoder-9B are averaged over five independent runs. A dash (—) means the score was not reported for that model.
* Terminal-Bench 2.1 (Terminus-2): We evaluate Terminal-Bench 2.1 using the Harbor/Terminus-2 framework with parser=json, temperature=1.0, top_p=1.0, and a 256K context window. Each run uses a 2-hour timeout with 32 CPU cores and 32GB RAM.
* Terminal-Bench 2.1 (Claude Code): We evaluate Terminal-Bench 2.1 using Claude Code 2.1.126 with parser=json, temperature=1.0, top_p=1.0.
* SWE-bench Verified and Pro: using the OpenHands harness with temp=1.0, top_p=0.95, 256K context window. Anti-hacking safeguards are applied throughout evaluation: Git history is removed from the local repository image to prevent access to prior solutions or commits; network access is disabled, preventing the model from retrieving external information or resources.
* NL2Repo: with temperature=1.0, top_p=1.0, 256K context, 48K output. Access to the specified GitHub repositories and pip packages is blocked to prevent reward hacking.
* HLE: Evaluated using GLM-5.3 as the judge model.
* MCP-Atlas: Evaluated using GLM-5.3 as the judge model.
* ClawEval: temp=0.6 and 256K context.
* Baseline scores for Ornith-1.5-9B, Ornith-1.0-9B, Qwen3.5-9B and Gemma-4-31B are taken from the Ornith-1.5-9B model card and were produced under that team's evaluation settings, which differ from ours on some benchmarks (context window, timeouts and judge model).
OxCoder-9B is developed by OrionLLM and released under the Apache 2.0 License.