Coder390's pitch to local coders: less wasted thinking

A community model aimed at the reasoning tail

Coder390 is attracting attention for a practical promise: spend less of a coding model’s output budget circling an answer it has already found. The community Qwen3.8-27B derivative appeared in this pass’s Hugging Face trend discovery, while an October 13 English X post discussed its shorter-reasoning approach and implications for local coding agents. The repository was created on October 4. Its October 14 update refreshed a quantization chart, so this is renewed discussion of an existing project rather than a new model release or evidence of measured growth.

The creator, nerkyor, says the model combines supervised fine-tuning with reinforcement learning that penalizes wrong, lengthy or empty outputs while preserving useful long reasoning. In the English card’s static-FP8 results, the author reports 90 correct LiveCodeBench answers out of 100, compared with 83 for the original FP8 baseline. Cases hitting the generation limit fall from 13 to three in that comparison. The stated protocol uses two RTX PRO 6000 GPUs, SGLang and a 94,208-token generation cap; these are the author’s results under that setup.

Why local users may want to inspect it

The repository offers GGUF builds for llama.cpp alongside other quantized formats, plus multi-token prediction and DFlash2 decoding options on supported engines. That makes it a candidate to try for users who already run local coding models and want to compare completed answers and wasted reasoning on their own tasks.

The card does not claim the long tail is eliminated: some coding problems still hit the cap, and results differ across precisions and runtimes. A reported FP8 score or decoding speed should not be transferred to every GGUF build or consumer GPU. The recent public discussion is evidence of interest, not an independent benchmark replication or an official Qwen endorsement.