Recursive releases Flint, a 40B-parameter open coding model that runs on laptops under Apache 2.0

Recursive, the AI startup led by former Salesforce chief scientist Richard Socher, released Flint on Wednesday, a free, open-weight coding model with 40 billion total parameters, of which only 2.5 billion are active at a time. The weights are available on Hugging Face under the Apache 2.0 license.

According to Recursive, Flint was built largely with the help of the company's automated research system, the Eureka Machine. The company says a small team built the model from scratch in under four months.

Specifications

• Architecture: Mixture-of-experts, 40B total parameters, 2.5B active per token
• License: Apache 2.0, free to download, use and modify
• Pricing: Free to download and run locally. Hosted API pricing not yet disclosed
• Context window: Not yet disclosed
• Training cutoff: Not yet disclosed
• Compression: Recursive says low-bit training shrank the model by 71% with almost no loss in quality

Speed claims

In Recursive's own tests, Flint generated about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip. The company says comparable models, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, managed 58 to 69 tokens per second on the same machine. On Nvidia's RTX Spark, Flint reportedly exceeded 150 tokens per second. These figures come from the company and have not been independently reproduced.

Benchmarks

Recursive claims Flint outperformed other models under 50 billion parameters on SWE-Bench Verified, which measures how well models fix real-world software bugs. Exact scores were not disclosed. On Recursive's own average across six tests, only Alibaba's Qwen3.8 27B scored higher than Flint, according to the company.

Recursive acknowledged that Flint needs more guidance than large paid coding assistants and recommends breaking large tasks into smaller steps. Wired's Will Knight ran Flint on a Mac with 48GB of memory and built several small programs, including a to-do app and a web research tool. He reported that Flint handled his requests correctly every time, unlike other small coding models he had tested, which hit errors or ran slowly compared with Claude or Codex.

The Eureka Machine

The Eureka Machine proposes ideas, runs experiments and uses the results to choose what to try next. Per Recursive's launch video, it contributed to pre-training, post-training and inference. The company credits it with a sparser model shape, the low-bit training approach and custom kernels. Recursive did not say how much researcher time the system saved. Humans oversaw the process.

Company background

Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia and AMD Ventures. In June it signed a $410 million compute deal with Amazon Web Services. Socher told Wired the long-term goal is full recursive self-improvement, with humans able to stop the process.

What this means

Flint's practical significance is its footprint. With 2.5B active parameters, it targets developers who want offline, private coding assistance without per-token fees. It sits at the opposite end of the scale from Meta's Muse Spark Open, which uses 816B parameters and needs at least 1TB of memory. The speed and benchmark claims are self-reported, and the missing context window and exact SWE-Bench Verified score limit direct comparison until independent evaluations appear.

The larger question is the development method. Recursive offers Flint as evidence that automated research can compress a from-scratch model build into under four months for a small team, but without figures on time saved, that claim cannot yet be measured. Whether each successive model makes the Eureka Machine a faster researcher, as Recursive says it should, will be tested by the company's next release.