Flint lands on Hugging Face: 40B MoE with 2.5B active params for on-device coding
pcuenq · x · 2026-10-22
Hugging Face's pcuenq welcomed Recursive's new Flint model, highlighting it as ideal for on-device use: a 40B MoE with 2.5B active parameters that runs fast on consumer GPUs, Macs and Sparks, claiming best-in-class coding and agentic workflow performance. Weights are available via transformers, llama.cpp and MLX.
More from Models
- One researcher pretrains coding model Mirendil-Base in 3.5 months for under $10M — bneyshabur · 2026-10-22
- OpenAI models now on Azure and AWS, coming soon to Google Cloud — romainhuet · 2026-10-22
- Flint cuts bike-demand validation RMSLE 30.8% in 3 hours of autonomous model iteration — cong_ml · 2026-10-22
- Running an autonomous research loop locally with Recursive Flint, from scheduling to bike-demand forecasting — cong_ml · 2026-10-22
- Developer calls out Grok's context bloat: no trimming, no window awareness — alexcovo_eth · 2026-10-22
- Grok's in-tweet chat takes ~20s to first token, dev says Google AI search is faster — BraceSproul · 2026-10-22