Новини

Recursive Launches Flint, a Laptop-Friendly Coding Model It Says Was Largely Built by AI

· TechCrunch

Recursive, an AI startup founded last year and led by former Salesforce chief scientist Richard Socher, has released its first model, Flint, a free coding model that runs locally on a laptop. According to TechCrunch, the company says the model was built largely with the help of its automated research system, the Eureka Machine, which let a small team produce it from scratch in under four months. Flint is presented as an early public test of recursive self-improvement, the contested idea that AI can build better AI, which then builds better AI still. Flint is a mixture-of-experts model with 40 billion total parameters, of which only 2.5 billion are active for any given task. That design keeps it light enough for ordinary hardware. It is far smaller than Meta's Muse Spark Open, which has 816 billion parameters and needs at least a terabyte of memory. Because it runs on the user's own machine, developers can work offline, keep code private, and avoid per-token cloud fees. It is available on Hugging Face under the Apache 2.0 license. Recursive reports strong speed numbers. On a MacBook Pro with an M4 Pro chip, Flint produced about 131 tokens per second, roughly double the 58 to 69 tokens per second of comparable models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark it exceeded 150 tokens per second. On SWE-Bench Verified, it beat other models under 50 billion parameters. Across Recursive's own six-test average, only Alibaba's Qwen3.8 27B scored higher. The company concedes that Flint needs more guidance than large paid coding assistants and recommends splitting big jobs into smaller steps. Wired's Will Knight, testing it on a 48GB Mac, found it handled his requests for small programs reliably, unlike many small coding models that stumble or run slowly compared with Claude or Codex. The Eureka Machine proposes ideas, runs experiments, and uses outcomes to choose what to try next. Recursive says it contributed across pretraining, post-training, and inference, including a sparser model shape, low-bit training that cut Flint's size by 71% with almost no quality loss, and custom kernels. The company did not quantify the time saved. Other labs are moving the same way: OpenAI said its AI agents logged 3.1 workdays for every human researcher workday by mid-August. Recursive emerged from stealth in May with $650 million at a $4.65 billion valuation, backed by GV, Nvidia, and AMD Ventures, and later signed a $410 million compute deal with Amazon Web Services. Socher told Wired the long-term aim is full recursive self-improvement under human supervision, with the ability to halt the process if needed. Flint's significance lies less in beating frontier systems than in showing that a lean, AI-assisted research pipeline can yield a competitive small open model quickly. If each generation improves the research system itself, the pace of development could accelerate, and local, free coding models may become a growing alternative to metered cloud assistants.

Recursive, an AI startup founded last year, has shipped its first model, Flint, and positions it as an early real-world test of recursive self-improvement. According to TechCrunch, Flint is a free coding model that runs on a laptop and was built largely with assistance from the company's automated research system, the Eureka Machine. A small team produced it from scratch in under four months.

What Flint is

Flint is a mixture-of-experts model with 40 billion total parameters, but only 2.5 billion are active at a time for any task. This sparse activation is what makes it practical on consumer hardware. By contrast, Meta's Muse Spark Open, an open-weight model aimed at Macs and released the prior week, has 816 billion parameters and requires at least a terabyte of memory.

Running locally brings practical benefits for developers: offline use, code that never leaves their machine, and no per-token charges of the kind cloud AI services bill. At a time when AI usage costs are a growing concern for teams, a capable free local model is an appealing proposition. Flint is downloadable from Hugging Face under the permissive Apache 2.0 license, allowing free use and modification.

Performance claims

Recursive's own testing shows Flint generating about 131 tokens per second on a MacBook Pro with an M4 Pro chip, about twice the 58 to 69 tokens per second similar models achieved on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, throughput exceeded 150 tokens per second.

On SWE-Bench Verified, which tests the ability to fix real software bugs, Flint outperformed other models below 50 billion parameters. Across a six-benchmark average compiled by Recursive itself, only Alibaba's Qwen3.8 27B ranked higher. These are vendor-reported figures, so independent replication will matter. The company is candid that Flint requires more hand-holding than big paid coding assistants and advises users to decompose large tasks into smaller steps.

Early outside impressions are favorable. Wired's Will Knight ran Flint on a Mac with 48GB of memory and had it build several small programs, such as a to-do app and a web research tool. He noted that small coding models typically hit errors or run sluggishly compared with Claude or Codex, yet Flint handled each of his requests successfully.

The Eureka Machine

The Eureka Machine generates hypotheses, runs experiments, and uses the results to choose its next steps. Recursive says it contributed to nearly all stages of Flint's creation, spanning pretraining, post-training, and inference. Named contributions include a sparser model architecture, low-bit training that shrank the model by 71% with almost no quality degradation, and custom kernels that accelerate core computations. The company did not disclose how much researcher time the system saved.

Recursive is not alone in automating research. OpenAI reported that by mid-August its agents were logging 3.1 workdays for each workday of a human researcher, suggesting that AI-driven research is becoming a mainstream tactic among leading labs.

Funding and ambitions

Recursive left stealth in May with $650 million in funding at a $4.65 billion valuation, with investors including GV, Nvidia, and AMD Ventures. CEO Richard Socher is Salesforce's former chief scientist. The company also struck a $410 million compute agreement with Amazon Web Services in June, which it described as being more about the number of agents than the number of employees.

The stated long-term goal, which Socher described to Wired, is full recursive self-improvement, with humans supervising and able to shut the process down if something goes wrong. The idea is that every new model improves the Eureka Machine as a researcher, which then builds the next model faster. Some researchers worry that self-improving AI could escape human control, but Recursive's present implementation is modest: a human-supervised research assistant that produced a laptop-scale model.

Analysis

Flint's importance is twofold. For developers, it adds a strong, permissively licensed local coding option that competes on speed and efficiency with offerings from Google, Nvidia, and Alibaba, and reduces dependence on metered cloud assistants such as Claude and Codex for simpler work. For the industry, it provides a concrete data point that AI-led research loops can compress development timelines for capable models. If the compounding effect Recursive predicts materializes, each release could handle longer autonomous work, raising both competitive pressure on rivals and the stakes of the oversight debate. Claims remain largely self-reported, so independent benchmarking and wider developer adoption will determine whether Flint becomes a staple of local coding workflows.

Посилання в статті