Qwen3.8-27B-Coder390 (nerkyor)

面向本地部署的多模态编程推理模型

入选理由: 多量化版本开箱即用,支持GGUF/vLLM/SGLang;基于Qwen3.8-27B做SFT+RLOO,增强代码和长推理,减少截断空答,适合本地部署。
对位: 对位 Qwen2.5-Coder-32B / Qwen3-32B
适合: 编程与推理任务本地部署 / 多精度量化与多模态评测
不适合: 低显存设备跑 BF16
规模: 27B · 100k · Q4 ~18GB / FP16 ~65GB
授权: Apache-2.0 · 需自查
框架: sglang / vllm / llama.cpp / ninfer
血统: 量化自 Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2
可信度: 8.8k 下载、124 赞；自报 GPQA 178/198、LCB 90/100

评分详情

作者: nerkyor
任务类型: image-text-to-text
推理库: 未指定
下载: 8,873
点赞: 124
许可证: Apache-2.0
标签: safetensors, gguf, qwen3.8, efficient-thinking, reasoning, coding, uncensored, sft, simpo, rloo, fp8, bf16, nvfp4, modelopt, quantization, calibration, static-quantization, compressed-tensors, gsq-rco, gsq, rco, hessian, imatrix, block128, mtp, dflash2, speculative-decoding, multimodal, sglang, vllm, llama.cpp, image-text-to-text, en, zh, base_model:nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2, base_model:quantized:nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2, license:apache-2.0, endpoints_compatible, region:us, conversational