Coder390 EfficientThink 27B

Coder390 is nerkyor’s community post-training project built on a Qwen3.8-27B EfficientThink base. The author describes alternating supervised fine-tuning and reinforcement learning to discourage reasoning that consumes the output budget without producing a final answer. Available formats include BF16, FP8, NVFP4, INT8, INT4 and GGUF; supported engines and evaluation results vary by package.

The card declares Apache-2.0 licensing and provides benchmark protocols, raw-result references and launch commands. Performance figures are the author’s reports. They should be read with the stated hardware, generation limits and precision, rather than treated as interchangeable results for every local build.