Artificial Intelligence
Recursive releases Flint from its AI research system
Recursive has published Flint, its first coding model, and said most of its development came from an internal system that proposed and tested ideas, while its main claims remain unverified.
- Recursive released Flint, its first language model, for programming tasks.
- The company said its internal system, Eureka Machine, helped with most stages of Flint's development.
- Recursive said Flint had 40 billion parameters, with 2.5 billion active per token, and that the weights were published on Hugging Face under Apache 2.0.
- Recursive said benchmark and speed claims had not been independently verified, and that the model was built for local use on Apple and Nvidia hardware.
Recursive released Flint, its first language model for programming work. The company said Eureka Machine had been involved in almost every stage of development, from proposing ideas to planning follow-up experiments.
Recursive said Flint used a mixture-of-experts design with 40 billion total parameters and 2.5 billion active per token. It said the model was built in four months and claimed it led several coding benchmarks among MoE models below 50 billion parameters, though those results had not been independently checked.
The company said Flint was intended for agentic coding on local hardware. It said the model produced more than 125 tokens a second on a MacBook Pro with M4 Pro and more than 150 on an Nvidia RTX Spark with 64 GB of memory.
Recursive said Flint's weights were available on Hugging Face under the Apache 2.0 licence, and that it supported several local runtimes. The company also released rec, its own inference engine for Apple Silicon.
Recursive, founded by Richard Socher and Tim Rocktäschel, came out of stealth in May 2026 and raised $650 million at a $4.65 billion valuation. The company said its longer-term aim was to build systems that improve their own development, while former Google DeepMind research lead Oriol Vinyals said AI-assisted self-improvement seemed plausible but an intelligence explosion was unlikely.
Named in this story
People
- Richard Socher
- Recursive co-founder
- Tim Rocktäschel
- former DeepMind researcher and Recursive co-founder
- Oriol Vinyals
- former Google DeepMind research lead quoted on self-improvement
Companies
- Recursive
- AI startup that released Flint and said Eureka Machine drove development
- Hugging Face
- platform hosting Flint's weights
- Apple
- hardware company whose chips Flint was optimised for
- Nvidia
- chip company whose hardware Flint was optimised for
- Anthropic
- AI lab cited as using AI in its own development
- OpenAI
- AI lab cited as deploying agents across its research pipeline
Products and systems
- Flint
- first language model for programming tasks and local coding
- Eureka Machine
- internal AI system that proposed, tested and planned experiments
- rec
- Recursive's inference engine for Apple Silicon
- MacBook Pro with M4 Pro
- device used in Recursive's speed claim
- Nvidia RTX Spark
- device used in Recursive's speed claim
- Claude
- Anthropic model said to handle much of its own development work
How the source tells it
The piece reads like a product launch, with company claims presented in detail and a late caution that the self-improvement idea is still unproven.
- vendor boosterism the model's performance and speed claims are relayed from the company before being marked as unverified
- speculative futurism the article frames recursive self-improvement and superintelligence as the company's long-term aim