Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh
Category: ai-ml
Tags: llm, quantization, inference-optimization, reasoning, qwen
Score: 6.7/10 (Innovation: 6, Technical: 7, Documentation: 8, Utility: 7)
Swift-Qwen3.8-27B is a reasoning-efficient fine-tune of Qwen3.8-27B that cuts thinking tokens by up to 58.3% while retaining near-identical accuracy, yielding up to 1.95x speedups. It targets cost and latency reduction for reasoning-heavy LLM deployments with drop-in compatibility across vLLM, SGLang, Transformers, and GGUF runtimes.
Target audience: AI/ML engineers and backend developers deploying reasoning LLMs
Repository: https://huggingface.co/ukisai/Swift-Qwen3.8-27b
View on Hacker News