ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF — new model trending #30 on Hugging Face
UkisAI published GSQ-RCO GGUF quantizations of Swift 1.5 Qwen3.8-27B, from 8.42 GB to 11.77 GB.
UkisAI published mixed-precision GGUF quantizations of Swift 1.5 Qwen3.8-27B, using per-tensor allocations from ISTA-DASLab's GSQ-RCO release. Standard files are IQ2_XS (8.42 GB, development KLD 0.189979), IQ2_S (9.26 GB, 0.134751), IQ3_XXS (10.09 GB, 0.097774), and IQ3_S (11.77 GB, 0.051265); optional MTP-head variants are about 0.35 GB larger. The card says Swift 1.5 uses 58.5% fewer thinking tokens and scores 0.35% higher than its base, for a 9.18× speed-up on several tasks; those figures are separate from the quantization KLD tests, which use a 512-token context and are not task-accuracy scores. All four refined files beat their matched Swift starting quants on seven held-out domains, but they are not uniformly better than the ISTA comparison quants.
- IQ2_XS through IQ3_S GGUFs range from 8.42 GB to 11.77 GB.
- Optional MTP files are about 0.35 GB larger and need a compatible runtime.
- KLD is versus Swift 1.5 BF16 on 512-token chunks, not task scores.
- Card claims 58.5% fewer thinking tokens and a 9.18× speed-up versus the base.
- llama.cpp with Qwen3.8 support is required; no vision projector is included.
- Organizations
- UkisAIISTA-DASLabHugging Face
- AI models
- Swift 1.5 Qwen3.8-27BQwen3.8-27B
From the original article
4 min read at huggingface.cotext-generation · 175,005 downloads · 160 likes · license other
By ukisai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.