Qwen3.8-Flash-Next · 1.7 bpw · supported on 48 GB+ Macs
35B MoE · supported on 16 GB+ Macs
The best desktop coworking app for local models.
Early access. Join the waitlist.
Blog
Research, releases, and the occasional field note on running models that never leave your machine.
PrimerOct 11, 202610 min
Scalar, vector, and trellis quantization — and a spherical trellis that pushes the perplexity-per-byte frontier while making decode cheaper.
Read the post→
All posts
Oct 11, 2026Primer