Recursive's Flint is an open-weights local coding model built in under four months

AIRecursive, through Sophia Yang's post, introduces Flint, its first model, a fast local coding model built with its Eureka Machine automated research system. The model is a 40B total, 2.5B active MoE with Apache 2.0 weights on Hugging Face, and it scores 54.9 on Terminal-Bench 2.1 versus 44.9 for Qwen3.6 35B-A3B. Dense Qwen3.8 27B still leads on coding with a Terminal-Bench score of 73.0, and Flint runs at 125+ tokens per second on an M4 Pro MacBook at short context.

Sophia Yang

@sophiamyang

View on X

TL;DR:

- Open-weights MoE coding model: 40B total, 2.5B active, Apache 2.0

- Built from scratch in under 4 months mostly by their automated research system, the Eureka Machine

- Main tricks: a sparser shape, the output layer quantized to 4 bits, and custom kernels, including a fused decode kernel on M4 and M5 Macs

- Beats MoE peers like Qwen3.6 35B-A3B on agentic coding: Terminal-Bench 2.1 54.9 vs 44.9, SWE-bench Verified 74.2 vs 73.4

- Dense Qwen3.8 27B still wins on coding, with Terminal-Bench 73.0 vs 54.9

- 125+ tok/s on an M4 Pro MacBook and 150+ on a 64GB RTX Spark at short context; about 69 tok/s on average across a 128K session

Richard Socher@RichardSocher

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

View quoted post on X

Source details

Sophia Yang · x.comPublished Oct 21, 1:15 PM ET · added here 1:27 PM ET

Original title: TL;DR:

Recursive's Flint is an open-weights local coding model built in under four months

Topics