| --- |
| license: apache-2.0 |
| base_model: openai/gpt-oss-120b |
| base_model_relation: finetune |
| library_name: mlx |
| tags: |
| - coding-agent |
| - agentic |
| - mlx |
| - gpt-oss |
| - codex |
| - apple-silicon |
| language: |
| - en |
| pipeline_tag: text-generation |
| --- |
| |
| # GPT-OSS Coder |
|
|
| A **gpt-oss-120b** version focused on improving practical coding-agent behavior in repository-level software engineering tasks. |
|
|
| > Also available in **[GGUF for llama.cpp / LM Studio / Ollama →](https://huggingface.co/exalandru/GPT-OSS-Coder-GGUF)** |
|
|
|
|
| # Version 1.1 |
| > This release adds three focused passes over re-curated data, each targeting a behaviour measured missing in the previous one: writing failure-injection regression tests before closing, iterating on multi-defect repositories past the first green suite, and honouring per-harness completion contracts. Every pass was gated by calibrated probes and a held-out benchmark ladder before being promoted. |
|
|
| ## How it works |
|
|
| It digs deeper into the repository, follows evidence to the root cause, and keeps iterating until the fix holds under real tests instead of stopping at a plausible-looking patch. |
|
|
| - Fixes the bug, not the symptom : traces the actual defect, not the first thing that looks broken |
| - Inspects more before editing, and re-runs tests after : more reads, more checks, fewer false successes |
| - Emits tool calls the harness can actually run : dramatically fewer rejected calls |
| - Revisits files when new evidence appears |
| - Reasons about state and invariants across components |
| - Continues iterating when the first implementation is incomplete |
| - Ends its turns with a real report of what was done — no empty summaries, no truncated turns |
| - Writes its own failing test before declaring victory : adds regression tests that fail on the unfixed code, then proves the fix against them |
| - Iterates past the first green suite : coupled defects don't survive the second pass |
| - Follows each harness's closing convention : a Codex task ends with a final report, a Cline task ends with a proper `attempt_completion` call |
| - Tightens what a fix requires : removes deprecated or unauthorized call forms instead of quietly keeping them accepted |
|
|
| The fine-tune also significantly reduced malformed JSON arguments. |
|
|
| Works even better with my **[Adversarial Agent Engineering](https://github.com/exalandru/Adversarial-Agent-Engineering)** pack of skills and rules |
|
|
| ## Runtime |
|
|
| <span style="color:orange;">Optimized for Codex.</span> |
|
|
| ### Codex GPT-OSS Server |
|
|
|  |
|
|
| It provides native Codex integration for GPT-OSS on MLX. Rather than exposing GPT-OSS only through a generic OpenAI-compatible compatibility layer, it is designed so that Codex can use GPT-OSS as a native local model while preserving the GPT-OSS/Codex protocol details. |
|
|
| This includes the native Codex Responses protocol, GPT-OSS Harmony handling, reasoning continuity across tool turns, and Codex-specific routing and metadata. |
|
|
| Agent loops work smoothly without the stalls and rejections you get with generic OpenAI-compatible endpoints. |
|
|
| **https://github.com/exalandru/Codex-GPT-OSS-Server** |
|
|
| ### Or run it directly with MLX |
|
|
| ```bash |
| mlx_lm.generate --model exalandru/GPT-OSS-Coder-MLX --prompt "Hello World!" |
| ``` |
|
|
| **Format:** MLX, MXFP4 experts + bf16 attention, ~61 GB on disk. Weights are consolidated — nothing to fuse or merge. Runs on Apple silicon with 96 GB unified memory; a full agent session peaks around 82 GB. |
|
|
| > **116.8B parameters** (MoE, ~5.1B active per token — which is why it decodes |
| > faster than much smaller dense models). The Hub badge shows ~24B because MLX |
| > packs eight 4-bit weights per uint32 element and the badge counts elements, |
| > not parameters. |
|
|
| --- |
|
|
| ## How it was trained |
|
|
| Supervised fine-tuning on ~10 000 steps carefully selected from real coding-agent sessions to isolate the targeted behavior : some of my personal sessions with Opus/Fable 5 and GPT 5.6 Sol, public SWE-agent, OpenHands, SWE-smith and Fable trajectories, keeping **only runs that actually resolved their issue**. A run that gave up, or ran out of context and submitted anyway, teaches exactly the habit this model is meant to shed, so those were filtered out. |
|
|
| Each training example is a real repository state plus the next action the successful agent took, so what is learned is the loop itself: look, edit, run, read the result, correct. |
|
|
| The fine-tune itself is deliberately small, a low-rank update on the last layers only, then consolidated back into the weights. The goal was to shift behaviour, not to overwrite what the base model already knows. |
|
|
| Custom small in-house benchmarks were used to validate the training. Models such as Qwen3.6, DeepSeek v4 Flash and other distilled gpt-oss variants all failed these benchmarks. Opus 5 and GPT 5.6 Sol served as references proving the tasks were solvable. |
|
|
| --- |
|
|
| *Built by [exalandru](https://github.com/exalandru). If you use it in a real |
| agent loop, the failure reports are more useful than the success ones.* |
|
|
