Persona programando en un computador portátil con código en pantalla

· Translated from the Spanish original

Artificial intelligence

Recursive releases Flint, a 40-billion-parameter coding model that runs on a laptop

Flint activates only 2.5 billion parameters per token and, according to its maker, was assembled in under four months with help from its automated research system.

What happened

  • Recursive, Richard Socher’s company, released Flint on October 21, 2026, its first model, aimed at agentic coding on local hardware.
  • It is a mixture-of-experts model with 40 billion parameters in total, of which it activates 2.5 billion per token.
  • Recursive says its “Eureka Machine,” an automated research system, helped develop almost the whole process: data, training recipes, infrastructure and inference optimization. A small team did it in under four months.
  • The company reports more than 125 tokens per second on a MacBook Pro with an M4 Pro and more than 150 on a 64 GB NVIDIA RTX Spark. According to TechCrunch, it is available on Hugging Face under an Apache 2.0 license.

Why it matters

  • A model that fits on a laptop changes the cost and privacy equation: the code never leaves the machine, it works offline and there is no charge per text processed.
  • For a Chilean software company with contracts that forbid sending code to foreign clouds, that is the difference between being able to use an assistant or not.
  • The comparisons are the company’s own: it says it leads mixture-of-experts models under 50 billion parameters on tests such as SWE-Bench Pro v2 and Terminal-Bench 2.1. Independent verification is missing.

The number

2.5 billion of 40 billion. Parameters Flint activates per token, the reason it runs on a laptop.

Context

Recursive came out of stealth in May with US$650 million in funding and a US$4.65 billion valuation, with GV, Nvidia and AMD Ventures among its backers, according to TechCrunch. Flint is presented as an early test of recursive self-improvement, the idea that AI builds better AI, with people supervising the process.

What’s next

  • Independent tests of the results the company published.
  • No dates for new Recursive models.

Bottom line

Open models are multiplying and competing on infrastructure: Xiaomi released MiMo-V2.6 with open weights, DeepSeek opened the engine that serves its models and AWS let PyTorch code run on its Trainium chips.

Sources


Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes