Swift 1.5 Qwen3.8-27B: GGUF variant built to stop overthinking

UkisAI released Swift 1.5 Qwen3.8-27B, a fine-tune of Qwen3.8-27B that uses 58.5% fewer mean thinking tokens on GPQA-Diamond with a slight score gain. GGUF quantizations range from 8.9 GB (IQ2_XXS) to 29.0 GB (Q8_0) for llama.cpp.

HackerNoon

Open source

Swift 1.5 Qwen3.8-27B: GGUF variant built to stop overthinking

GPQA-Diamond: 88.59% vs 88.28% for the base, mean thinking tokens down from 15,014 to 8,717
LiveCodeBench v6: 81.71% vs 76.76%, using 24.5% fewer mean reasoning tokens
Terminal-Bench 2.1: 72.13% vs 69.21%, agent-call tokens down from 52,265 to 43,733
GGUF tiers span 8.9 GB to 29.0 GB; up to 9.18x speed-up reported on several tasks

#Qwen#UkisAI#Llama.cpp

Swift 1.5 Qwen3.8-27B: GGUF variant built to stop overthinking

Swift-Qwen3.8-27B Cuts Reasoning Tokens Without Sacrificing Much Accuracy

Qwen3.8-27B on one RTX 3090 vs two: +20% decode, +14% cold prefill

Qwen3.8-Flash-Next vs Qwen3.8-27B: 125B Parameters, 6B Active — What the Qwen4 Preview Changes

DeepMind releases AlphaGenome Atlas: 9 billion DNA variants made public