hn · Show HN · 09-16 14:24 51 分 ← 返回列表

Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh

Hi everybody, we post-trained Qwen 3.8 27B to be more efficient by figuring out which tokens were linked to overthinking and penalizing them without "attacking" the reasoning length directly then fixed the accuracy with a bit of secret sauce (hint On-Policy Distillation) and achieved great results (-58% length, 1.95x speed up, <1% accuracy loss) so we wanted to open-source it and hear the feedback of the community. This is the link to the model: https://huggingface.co/ukisai/Swift-Qwen3.8-27b

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：大模型推理时过度思考、反复推理导致 token 消耗高、延迟大、API 成本上升。

目标用户：本地/自托管大模型用户、AI 应用开发者、对推理成本和延迟敏感的团队。

证据：It's at 80k downloads in 3 days with independent evals here; We are also providing a Free Research Purpose API (OpenAI compatible), courtesy of Nvidia... It's limited at 5RPM; -58% length, 1.95x speed up, <1% accuracy loss

变现：按 token 计费的托管推理 API；企业私有化部署与微调服务；面向 LLM 应用的省钱/降延迟代理服务。

MVP：用 vLLM/llama.cpp 部署 Swift-Qwen3.8-27B GGUF，提供 OpenAI 兼容 API，免费 5 RPM，付费提高速率；或做一个 token 节省代理，对比原始 Qwen 并展示成本节省。

竞品：Qwen3.8-27B、DeepSeek-R1 distills、OpenAI o-series、Together AI/Fireworks 推理托管、LocalLLaMA 社区量化版本 (Bartowski, MLX)

摘要：Show HN 发布 Swift-Qwen3.8-27B，声称通过后训练减少 58% 思考长度、推理加速 1.95 倍，精度损失 <1%，3 天获 80k 下载，并提供免费研究 API。社区已产出 GGUF/MLX/NVFP4 等多种量化版本。机会在于降低推理成本与延迟的托管 API 或工具，但缺少明确付费数据，且核心训练需较大算力。