UkisAI compresses Swift 1.5 27B into an 8.42GB GGUF

UkisAI has released four mixed-precision GGUF versions of Swift 1.5, a post-trained Qwen3.8-27B reasoning model. The smallest model file occupies 8.42GB, lowering the storage and memory requirements for local inference on consumer GPUs and unified-memory Macs.

The release combines per-tensor bit allocations from ISTA-DASLab’s quantization research with a Swift-specific calibration and refinement process. UkisAI reports that the resulting files track the BF16 model’s token probabilities more closely than its initial Swift quantizations across prose, code, mathematics and multilingual test data.

Two sets of efficiency claims

Swift 1.5 itself provides the first efficiency gain. According to UkisAI, the full-precision post-trained model uses 58.5% fewer reasoning tokens than its Qwen base while scoring 0.35% higher across the cited tasks, producing a reported 9.18× speed-up on several evaluations.

Those figures apply to the full-precision Swift 1.5 model. The GGUF release adds weight compression, but the model card does not report equivalent downstream task scores or end-to-end speed measurements for every quantized tier. Its quantization results focus primarily on distribution similarity.

Four files trade size for fidelity

Each tier assigns different bit widths to individual tensors according to their sensitivity. The reported file sizes and development-set Kullback-Leibler divergence, or KLD, are: