Recursive Releases Flint, Free 40B-Parameter Laptop Coding Model Built Largely With AI

4.0 · 10d ago · US · AI

The Facts

  • AI startup Recursive released its first model, Flint, on Wednesday as a free coding model that runs on a laptop. TechCrunch only
  • Flint is available to download on Hugging Face under the Apache 2.0 open-source license. TechCrunch only
  • Recursive says its automated research system, the Eureka Machine, helped with nearly every step of development, from pre-training to post-training and inference, letting a small team build Flint from scratch in under four months. TechCrunch only
  • Flint is a mixture-of-experts model with 40 billion parameters, of which 2.5 billion are active at a time. TechCrunch only
  • In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with M4 Pro chip, versus 58 to 69 for similar models including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, and over 150 tokens per second on Nvidia's RTX Spark. TechCrunch only
Dive deeper · 6 more factsShow less
  • On SWE-Bench Verified, Flint outperformed other models under 50 billion parameters; on Recursive's own six-test average, only Alibaba's Qwen3.8 27B scored higher. TechCrunch only
  • Recursive says Eureka's contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels. TechCrunch only
  • Wired's Will Knight reportedly used Flint on a 48GB Mac to build several small programs and said it handled his coding requests every time. TechCrunch only
  • Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, from backers including GV, Nvidia and AMD Ventures; CEO is Richard Socher, formerly Salesforce's chief scientist. TechCrunch only
  • Recursive signed a $410 million compute deal with Amazon Web Services in June. TechCrunch only
  • The company acknowledged Flint needs more guidance than large paid coding assistants and suggested splitting large tasks into smaller steps. TechCrunch only

Context

Recursive, an AI startup founded last year and focused on building AI that can improve itself, released Flint, its first model. The company describes it as an early public test of recursive self-improvement, in which an AI research system overseen by people helps build the next model.

Because Flint runs locally, developers can use it offline, keep code private and avoid cloud usage fees. Performance figures come from Recursive's own testing. CEO Richard Socher told Wired the long-term goal is full recursive self-improvement with human oversight and the ability to stop the process.

How the Agent Panel Scored This

Independent AI personas each scored importance 1–10; the story's score is a trimmed mean (highest and lowest votes dropped). Vote spread: 1.

4 Critical Reader A free open-weight coding model release from a startup is a concrete product launch with downloadable impact, but its performance claims are self-reported, the user base is niche (developers), and it is a single independent-source story, so it sits in the minor-news range.
4 Policy Expert A startup releasing a free open-weight coding model is a concrete product launch of niche-to-moderate interest, with performance claims mostly self-reported and no governmental or broad societal impact.
4 News Editor A concrete product release of a free, open-weight coding model by a well-funded startup matters to developers and the AI niche, but its reach is limited to technical users and the performance claims are self-reported, so it is a moderate-to-minor story.
4 Business Analyst A startup releasing a free, open-weight coding model is a notable product launch with self-reported benchmarks and a single source, affecting developers more than broad markets or economy.
4 Science & Tech Analyst A free, open-weight 40B MoE coding model released by a funded startup is a notable AI release, but its performance claims are self-reported and it is a single-company product launch with limited broad impact.
4 Global Affairs Analyst A startup's release of a free open-weight coding model is a notable AI product launch with mostly self-reported performance claims, but it has limited geopolitical or broad societal impact.
3 Informed Citizen A startup releasing a free open-weight coding model is a concrete event with modest impact mainly on developers, and its performance claims are self-reported, so it is minor news for ordinary people.

Source

← All stories