Tess-4-27B - a huginnfork Collection

huginnfork 's Collections

ThinkingCap-Qwen3.6-27B

Tess-4-27B

updated 9 days ago

FP8 & NVFP4 quants of migtissera/Tess-4-27B, a Qwen3.6-27B reasoning fine-tune. KLD is measured against bf16 Tess, so it isolates quant loss.

huginnfork/Tess-4-27B-FP8

Image-Text-to-Text • 28B • Updated 19 days ago • 302 • 1

Note FP8 attnbf16 — MLPs only, the whole self_attn path stays bf16. KLD 0.0078 nats vs bf16 Tess, the lowest of any build we ship. 75.2 % vLLM MTP draft acceptance.
huginnfork/Tess-4-27B-NVFP4A16

Image-Text-to-Text • 20B • Updated Jul 8 • 14

Note NVFP4 W4A16, ~28.6 GiB. KLD 0.0418 nats, PPL 7.056 (+3.49 %) vs bf16 Tess.