Recursive unveils Flint coding model designed to run on laptops
Recursive has released Flint, its first public AI model: a free coding model designed to run on a laptop. The startup says its automated research system, the Eureka Machine, contributed to nearly every stage of Flint’s creation, enabling a small team to build the model from scratch in under four months.
Flint is a mixture-of-experts model with 40 billion parameters, but it activates only 2.5 billion parameters for a given task. Recursive says that design makes the model light enough for an ordinary laptop while retaining coding capability. The model can be downloaded from Hugging Face under the Apache 2.0 licence, which permits free use and modification.
Local coding model and performance claims
Running on a user’s own computer means Flint can be used offline, keeps code on the device and avoids the usage-based fees attached to cloud AI services. Recursive reported throughput of about 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip, compared with 58 to 69 tokens per second for similar models on the same hardware, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning.
On Nvidia’s RTX Spark, Recursive said Flint exceeded 150 tokens per second. The company also said Flint outperformed other models below 50 billion parameters on SWE-Bench Verified, a benchmark for fixing real-world software bugs. Across six tests averaged by Recursive, it said only Alibaba’s Qwen3.8 27B scored higher.
The company cautioned that Flint needs more guidance than large paid coding assistants and recommended splitting substantial assignments into smaller steps. In an early hands-on test, Wired’s Will Knight used the model on a Mac with 48GB of memory for small programs including a to-do list application and a web research tool, reporting that Flint completed his coding requests.
Eureka Machine is central to Recursive’s strategy
Recursive describes the Eureka Machine as a research system that generates ideas, runs experiments and uses results to determine subsequent experiments. Its launch video says the system contributed from pre-training through post-training and inference. The company attributes a sparser model shape, custom computational kernels and low-bit training to the system; it says the latter reduced Flint’s size by 71% with almost no quality loss.
The release is an early public test of recursive self-improvement: AI helping build better AI that can then improve future systems. Recursive characterises its present approach more narrowly as a human-supervised research assistant, rather than an autonomous process outside human control. CEO Richard Socher told Wired that the longer-term objective is full recursive self-improvement with human oversight and the ability to stop the process if problems arise.
That ambition is tied to the company’s compute plans. In June, the company described Recursive’s AWS compute agreement for AI agents as an arrangement focused less on headcount than on agent count, signalling the infrastructure required if automated research systems take on longer stretches of work. Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, with backers including GV, Nvidia and AMD Ventures.
What software teams can assess now
For businesses, Flint offers a concrete local-model option rather than a claim that coding work can be fully automated. Teams can assess it on bounded development tasks where offline use, code privacy and predictable costs matter, while preserving human review for task definition, testing and deployment decisions.