UkisAI Shrinks Swift 1.5 27B Model to 8.42GB

UkisAI has released mixed-precision GGUF quantizations for Swift 1.5 Qwen3.8-27B, a post-trained 27-billion-parameter reasoning model. By combining per-tensor bit allocations based on GSQ and RCO research from ISTA-DASLab with Swift-specific refinement, the team compressed the model into file sizes ranging from 8.42GB to 11.77GB. This reduction lowers the memory threshold required to run local inference on consumer GPUs and unified-memory Macs.

The core Swift 1.5 architecture delivers substantial operational improvements over its base model. Full-precision Swift 1.5 consumes 58.5% fewer thinking tokens while outperforming its Qwen base by 0.35%, delivering a reported 9.18× speed-up on multiple evaluations. UkisAI structured the GGUF releases into four precision tiers: IQ2_XS, IQ2_S, IQ3_XXS, and IQ3_S. Each tier assigns bit widths based on tensor sensitivity and includes optional variants featuring multi-token prediction heads.

While UkisAI noted that the quantization metrics primarily measure token probability alignment through Kullback-Leibler divergence rather than explicit task accuracy for every tier, the builds track full-precision output closely across code, prose, mathematics, and multilingual tests. This per-tensor allocation preserves model behavior while dramatically lowering VRAM requirements.

For machine learning practitioners, these builds make deploying complex 27B reasoning models locally far more feasible. The quantized files offer immediate compatibility with tools like llama.cpp, Ollama, LM Studio, vLLM, and Docker Model Runner. Enterprise teams should note that UkisAI distributes the release under the Swift Open License v1.0 rather than Apache 2.0, necessitating legal review prior to commercial adoption.