Recursive launches Flint, a laptop coding model it says was largely built by AI

Recursive, an AI startup founded past year, is starting to unrecorded up to its name. The company, which is trying to physique AI that tin amended itself, released its archetypal model, Flint, connected Wednesday. Flint is simply a escaped coding exemplary that tin tally connected a laptop and was built mostly with the assistance of AI.

Recursive’s automated probe system, which it calls the Eureka Machine, helped with astir each measurement of Flint’s development, the institution says. That fto a tiny squad physique the exemplary from scratch successful nether 4 months.

Flint is an aboriginal nationalist trial of recursive self-improvement: the thought that AI tin physique amended AI, which successful crook builds adjacent amended AI. It’s besides 1 of the astir debated ideas successful the industry. Some researchers fearfulness that AI improving itself could gaffe retired of quality control. For now, Recursive’s mentation is much limited: an AI probe adjunct overseen by radical that produced a exemplary tiny capable for a laptop.

The institution emerged from stealth successful May with $650 cardinal successful funding astatine a $4.65 cardinal valuation, from backers including GV, Nvidia, and AMD Ventures. Recursive is led by CEO Richard Socher, Salesforce’s erstwhile main scientist.

Because Flint runs connected a user’s ain machine alternatively than a company’s unreality servers, developers tin usage it offline, support their codification private, and debar the usage fees that AI unreality services complaint based connected however overmuch substance the exemplary processes. The exemplary is escaped to download connected Hugging Face, the fashionable tract for sharing AI models, nether the Apache 2.0 license, an open-source licence that allows escaped usage and modification. 

Flint is simply a mixture-of-experts model, meaning it uses lone a tiny portion of itself for immoderate fixed task: conscionable 2.5 cardinal of its 40 cardinal settings, known arsenic parameters, astatine a time. That makes it airy capable for an mean laptop and acold smaller than Meta’s Muse Spark Open, an unfastened exemplary built for Macs that was released past week. Muse Spark Open uses 816 cardinal parameters and needs astatine slightest a terabyte of memory. 

Flint is besides fast. In Recursive’s tests, it generated astir 131 tokens (chunks of text) per 2nd connected a MacBook Pro with Apple’s M4 Pro spot — astir doubly the 58 to 69 tokens per 2nd that akin models managed connected the aforesaid machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, Flint topped 150 tokens per second.

On SWE-Bench Verified, a benchmark that measures however good AI tin hole real-world bundle bugs, Flint outperformed different models nether 50 cardinal parameters. On Recursive’s ain mean crossed six tests, lone 1 model, Alibaba’s Qwen3.8 27B, scored higher than Flint.

The institution acknowledged that Flint would request much guidance than the large paid coding assistants, and suggested breaking ample tasks into smaller steps. Still, aboriginal reviews were positive.  Wired’s Will Knight used Flint connected a Mac with 48GB of representation to physique respective tiny programs, including a to-do database app and a instrumentality that scours the web for research. He recovered that smaller coding models often ran into errors oregon were dilatory compared to Claude oregon Codex, but Flint nailed his coding requests each time. 

The Eureka Machine comes up with ideas, runs experiments and uses the results to determine what to effort next. According to the company’s launch video, it contributed to astir each portion of gathering Flint, from pre-training to post-training and inference. 

Recursive said that Eureka’s contributions see a sparser exemplary shape, low-bit grooming that shrank Flint by 71% with astir nary nonaccomplishment successful prime and customized kernels, which are specialized programs that velocity up the underlying computations.

Recursive didn’t accidental however overmuch clip the Eureka Machine saved its researchers, but it’s not unsocial successful leaning connected AI to bash research. OpenAI said last month that by mid-August, its AI agents were logging 3.1 workdays for each 1 worked by a quality researcher.

If Recursive’s stake pays off, each caller exemplary should beryllium capable to instrumentality connected longer stretches of enactment without quality help. The institution antecedently told TechCrunch that its $410 cardinal compute woody with Amazon Web Services successful June was “less astir headcount and much astir cause count.”

The longer-term goal, Socher told Wired, is for the Eureka Machine to execute afloat recursive self-improvement, with humans overseeing the process and capable to halt it if thing goes wrong. Recursive says each caller exemplary should marque the Eureka Machine a amended researcher, helping it physique the adjacent 1 faster.