Nace ships 71.8GB DeepSeek V4.1 Flash coding quant with 39% lower KL than rivals
m4rkGS · reddit · 2026-11-19
Open-source lab nace.ai released two GGUF quants built with its Weight Space Quantization method:
- Hy4-flash general quant: 107.49GB (KL 0.233) and 127.22GB (KL 0.133), 2-3x lower KL than comparable Xylolabs, AngelSlim, and Atomic Chat files.
- DeepSeek V4.1 Flash Coding: a specialized 71.8GB quant of the 763B model, calibrated on real coding work (agent sessions, contest solutions, commits), keeping the 200 of 384 experts per layer that code routes to most (90% of code routing), with per-layer mixed precision. Held-out KL is 0.140 vs 0.231 for ISTA Coder at the same size (39% lower), and it solves 37/38 LiveCodeBench v6 problems on both seeds.
- Runs on a 48GB M5 Pro Mac via llama.cpp; MIT-licensed with weights and Engram table published separately.
More from coding & agent
- AI is letting hackers attack at scale, warns security pro amid week of outages — IanAndrewsDC · 2026-11-19
- Dev finds hacky way to gate agent flows with verifiable human authorization — csuwildcat · 2026-11-19
- Code is a liability, not an asset: AI coding makes restraint more valuable — bendee983 · 2026-11-19
- Gemini 4 Neon's 39% SWE-sweep questioned: results hinge on the harness — OfirPress · 2026-11-19
- Why the most AI-pilled engineers cluster at labs: unlimited tokens, agents, and humans in the loop — dbreunig · 2026-11-19
- Harness is a hidden variable: researcher blasts AI labs as terrible engineers — gerardsans · 2026-11-19