Recursive Luxury Founding Team's First Submission: Letting AI Participate in Building AI, Training a 40B Model in 4 Months

Beating AI News Flash: Recursive, which secured $650 million in funding this May at a valuation of $4.65 billion, has finally delivered its first model, Flint.

It is specifically designed for local coding agents, with approximately 40 billion total parameters and only 2.5 billion activated at a time. The team developed this model from scratch in less than 4 months.

Recursive's 8 co-founders are all heavy hitters:

1. Alexey Dosovitskiy: First author of Vision Transformer (ViT), driving Transformer's entry into computer vision.

2. Yuandong Tian: Former Meta FAIR Research Director, long engaged in reinforcement learning and large model reasoning research.

3. Tim Rocktäschel: Former Head of Open-Ended Learning research at Google DeepMind, Professor at UCL.

4. Jeff Clune: Author of the "Darwin Gödel Machine" paper, researching how to enable AI to modify its own code and continuously improve.

5. Josh Tobin: Previously led OpenAI's Codex and Deep Research teams.

6. Tim Shi: Former OpenAI researcher, co-founder of enterprise AI unicorn Cresta.

7. Richard Socher: Recursive CEO, former Salesforce Chief Scientist, founder of You.com.

8. Caiming Xiong: Former Senior Vice President of Salesforce AI Research.

The team developed an automated research system called Eureka Machine. It can autonomously propose improvement plans, write code, run experiments, and continue adjusting based on results. Flint's model architecture, training methods, quantization, and inference optimization all incorporated research outputs from this system.

Recursive's published tests show that Flint scored 74.2% on SWE-bench Verified. Paired with a dedicated inference engine, on an M4 Pro MacBook Pro, short-context generation speed exceeds 125 tokens per second.

The model weights have been released under the Apache 2.0 license.