Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Nace releases open-source KQuant MLX packages for Qwen models

AI summary

Nace released free, open-source KQuant packages in standard MLX format for Qwen3.5-4B, Qwen3.5-9B, Qwen3.8-27B, Qwen3.6-35B-A3B, and Qwen3.8-Flash-Next. In the developers’ measurements, the packages had KL divergence from full-precision next-token predictions up to 3.7 times lower than cited public 4-bit baselines. On an M5 Pro, selected models were reported to write 6%–15% faster with Kapitsa 0.2.35 and the KFlash drafter, although baseline tests used earlier Kapitsa versions. The packages require no custom kernels or special loader. GGUF versions were still being tested ahead of a planned public release.

In-depth insight

The release provides standard MLX builds for local Qwen use on Macs, alongside the developers’ reported fidelity and selected write-speed results. The KL measurements show similarity to full-precision predictions and are not task-accuracy scores.

#Nace#KQuant#Qwen#MLX#Hugging Face#Kapitsa#KFlash#Qwen3.5-4B#Qwen3.5-9B#Qwen3.8-27B#Qwen3.6-35B-A3B#Qwen3.8-Flash-Next

Original · 2 min readRead the original
Group chat

No comments yet. Start the conversation.