Coder390 Trains Qwen3.8-27B to Stop Wasting Tokens Before Failing
A community post-trained Qwen variant uses RLOO to stop endless re-derivation, cutting 94K-token truncations and pushing LiveCodeBench from 83 to 90.
Takeaways
- Coder390 is a community post-training of Qwen3.8-27B targeting unproductive reasoning tails.
- LiveCodeBench jumped from 83 to 90, with 94K truncations dropping from 13 to 3.
- GPQA hits 178/198 and MMLU 450/500 on static FP8, matching or beating the base.
- Uses RLOO with asymmetric rewards: wrong-and-long is penalized, long-and-correct is not.
- DFlash2 speculative decoding reaches 180 tok/s single-request, 1,057 tok/s aggregate at C16.
- Ships BF16, FP8, three NVFP4 tiers, INT8, INT4, and multiple GGUF variants under Apache-2.0.
A new community post-training release, Coder390 model card, targets a costly failure mode in long-context reasoning. A model can produce a valid solution early, continue revising it with loops such as “Wait” and “Actually,” exhaust its context window, and return no final answer. Coder390 packages five Safetensors format families, GGUF and NInfer builds, and two speculative-decoding options. Its Hugging Face page reported more than 56,000 downloads at publication.
Why the model runs out the clock
Coder390 focuses on traces that contain a workable answer before repeated self-correction consumes the remaining token budget. Because the solution already appears in many failed traces, the training objective targets stopping behavior and final-answer production.
The authors compared static FP8 Coder390 with the original Qwen3.8-27B FP8 checkpoint using the same 100K context and a 94,208-token generation cap. GPQA measures graduate-level science reasoning, MMLU covers broad academic knowledge, and LiveCodeBench, abbreviated LCB, evaluates coding.
| Benchmark | Original FP8 | Coder390 FP8 | Original cap hits | Coder390 cap hits |
|---|---|---|---|---|
| GPQA | 177/198 | 178/198 | 4 | 1 |
| MMLU | 444/500 | 450/500 | N/A | N/A |
| LiveCodeBench | 83/100 | 90/100 | 13 | 3 |
The largest behavioral change appears in completion rates. Coder390 reduced GPQA cap hits from four to one and LiveCodeBench cap hits from 13 to three, allowing more runs to return an answer before the server terminates generation.
Reward the answer, penalize the loop
The training pipeline alternates supervised fine-tuning with RLOO, short for REINFORCE Leave-One-Out. Supervised examples establish target behavior, while RLOO estimates the relative value of sampled responses without the separate critic, or value model, required by PPO.