Flint, an open-weight coding model, was released by Recursive on October 21, 2026: 40 billion parameters in total with only about 2.5 billion active at a time, light enough to run directly on a laptop, with weights free on Hugging Face under the Apache 2.0 license. What draws attention is not the parameter count but the model's origin. Recursive says its automated research system, the Eureka Machine, took part in nearly every step from pre-training to inference optimization, and a small team built the model from scratch in under four months.
Small frame, roughly twice the speed of its peers
Flint uses a mixture-of-experts architecture, and its model card lists a 262K-token context window. In Recursive's own tests, on a MacBook Pro with an M4 Pro chip it generated about 131 tokens per second, against 58 to 69 for comparable models on the same machine, roughly twice as fast; on Nvidia's RTX Spark it topped 150. The scorecard is likewise the company's own: 74.2% on SWE-bench Verified, leading mixture-of-experts models under 50 billion total parameters, and 54.9% on Terminal-Bench 2.1. Wired's hands-on testing backs the pitch: on a Mac with 48GB of memory, Flint built a to-do list app, a research-gathering tool and a small website, getting each request right. Running locally means code never leaves the machine and no per-token fees accrue; Nvidia refreshed its whole local Spark hardware line this year, betting on exactly this move of large models back onto the desk.
The Eureka Machine: AI helping to build AI
The Eureka Machine proposes ideas, writes the code to test them, reads the results and decides what to try next, running many experiments in parallel. Its credited contributions to Flint include a sparser model shape, low-bit training that shrank the model by 71% with almost no loss in quality, and custom kernels tuned for Apple silicon and RTX Spark. Recursive was founded by Richard Socher, Salesforce's former chief scientist; when it emerged from stealth in May it had raised $650 million at a $4.65 billion valuation, with backers including GV, Nvidia and AMD Ventures.
Do not hear AI building AI as out of control
Three buckets of cold water are in order. First, the benchmarks and speeds above are Recursive's own measurements, and independent replication has not caught up; even on the company's six-test average, only Alibaba's Qwen3.8 27B sits ahead of Flint, so this is no clean sweep. Second, Recursive itself concedes that on complex jobs Flint needs more guidance than the large paid coding assistants, and works better when big tasks are broken into smaller steps. Third, recursive self-improvement here is still an AI research assistant under human oversight, and Socher states plainly that the system remains under human control and can be stopped. Flint's value is that it turns how far AI can participate in building AI from an argument into a downloadable sample anyone can test on the spot.
