Project · October 2026

Katy

A quantization of Swift-1.5-Qwen3.8-27B for llama.cpp and the FLE engine. As with Rhea: this page states only what has been measured.

Specs

Base model	Swift-1.5-Qwen3.8-27B (ukisai), derived from Qwen3.8-27B
Quant	IQ4_XS with imatrix (production), file 15.3 GB
Quality against BF16	KL divergence vs BF16 (E, W, K): IQ4_XS 0.0645 / 0.0127 / 0.0160; top-1 92.5 / 95.9 / 95.0 %. ARC-Challenge 96.3 to 96.5 %, GSM8K 96.8 to 97.2 %. Source: model card on the Hub
Decode speed	127 tok/s total on the mixed task set (DFlash2 draft head + ngram-mod ng8, ABAB, drift 0.5 %). About 80 tok/s on fixed tasks, 46.6 tok/s without speculation. Measured on an RTX PRO 4500 Blackwell (32 GB)
License	Swift Open License v1.0, derived work marked as such (NOTICE). Terms reviewed, publication approved (Max, 22 Oct 2026)
Files	One Hugging Face repo: sirfyyn/Katy-27B-GGUF, public. IQ2_S, IQ3_S and IQ4_XS with imatrix, plus the DFlash2 draft head as Q8_0 and the imatrix; checksums in SHA256SUMS