$650M-Backed Recursive Says AI Helped Build Its First Model

Recursive, the AI lab run by Richard Socher, released Flint on 21 October 2026, its first model: a fast coding model that runs on a laptop and that the company says it built with its own automated research system. “Our small team of people and agents built it from scratch in under four months,” Recursive said in its launch post.

That system, which Recursive calls the Eureka Machine, “helped develop nearly every part of Flint’s model-building stack, from data and training recipes to infrastructure and inference optimization,” the company said. Recursive came out of stealth in May with $650 million in funding, TechCrunch reported at the time.

Business Pill · OPEN WEIGHTS

A short Business Engineer explainer on open weights: renting a model through an API versus running it yourself. It teaches the general idea only and says nothing about Recursive or Flint.

What Recursive Released

Flint is an open-weight model: Recursive publishes the trained weights, so anyone can download the model and run it on their own computer instead of paying per request to a hosted service. The weights are on Hugging Face under the Apache 2.0 license, according to the model card.

The model card describes Flint as a mixture-of-experts model “with 39.8 billion total parameters and approximately 2.5 billion active per token.” Only a small slice of the model works on each word it produces, which is what makes it fast on modest hardware.

Recursive reports that Flint “scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.” Those are the company’s own results: the card says “† marks scores measured by Recursive; other comparison scores are developer-reported.”

What Was Posted on X

Richard Socher, who leads Recursive, posted the launch on X at 16:29 UTC on 21 October 2026, calling Flint the lab’s first model and a fast local coding model built with its Eureka Machine. His post claims 125+ tokens per second on a MacBook Pro and 150+ on RTX Spark, a build from scratch in under four months, and open weights on Hugging Face. These are Recursive’s own figures, which this piece reports from the launch post and model card.

How Fast It Runs on a Laptop

Recursive says that with its inference engine, “Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts.” A token is roughly a word fragment.

Socher put the pitch in one line on X: “Private, unmetered intelligence on your own machine.” Recursive says Flint works with tools including llama.cpp, Ollama, LM Studio and vLLM, and it is also releasing rec, a lightweight inference engine for Apple’s M4 Pro and M5 Pro and Max chips.

The launch post adds a caveat for users: “While Flint is very capable, it may need more guidance than the larger closed-source coding models you’re used to.”

What the Eureka Machine Did

Recursive describes the Eureka Machine as a system that “automates the full research loop: proposing ideas, implementing them, testing them, and using the results to decide what to try next.” WIRED’s report, as summarised by Techmeme, says the system “automated most of the development.”

The launch post gives three examples. Starting from a baseline with 35 billion total and 3 billion active parameters, the system explored expert configurations and arrived at Flint’s 40-billion, 2.5-billion-active design. It then searched for ways to compress the model, and finally wrote specialised code for Apple Silicon and NVIDIA RTX Spark, including what Recursive calls “a Metal megakernel that fuses the entire decode forward pass.”

As we read it, AI building AI here means an automated research assistant working alongside a small human team, not a model training itself unsupervised. Recursive’s own wording is that the system “contributed throughout Flint’s development.”

Who Is Behind Recursive

TechCrunch described Socher in May as “best known for founding the early chatbot startup You.com and, before that, his work on ImageNet.” It reported that Peter Norvig and Cresta co-founder Tim Shi are among the researchers working with him.

Recursive states its goal plainly in the launch post: “Our goal at Recursive is to build recursively self-improving superintelligence to automate knowledge discovery.” It calls Flint “our starting point.”

Connecting the Coverage

We first covered Recursive in September, when Socher described a simulated economy on The MAD Podcast (our earlier Recursive piece). Flint is the lab’s first released model.

Local AI has been a theme this week. On 20 October, Perplexity said its Computer agent can now run fully on Macs with Apple silicon (our Perplexity local piece). At GTC Berlin on 21 October, Nvidia’s Jensen Huang held up an RTX Spark laptop, the hardware Recursive cites for its 150-token speed (our GTC Berlin wrap).

This week, Anthropic, OpenAI and Google each said their AI systems helped find software flaws (our AI bug-finding piece). Recursive’s claim is a related one: AI doing the research, not just the coding.

The Structural Read

As we read it, the launch showcases a research process as much as a model: Recursive says the Eureka Machine “automates the full research loop: proposing ideas, implementing them, testing them, and using the results to decide what to try next.”

As we read it, local speed is the product: Recursive says the system “produced specialized kernels for low-bit projections, MoE operations, attention, and prompt processing, tailored to the model and hardware.”

As we read it, Recursive frames Flint as a first step: “Our goal at Recursive is to build recursively self-improving superintelligence to automate knowledge discovery.” It adds: “Flint is our starting point.”

Three Implications

FOR DEVELOPERS The model card lists the Apache 2.0 license, and Recursive says Flint works with tools including llama.cpp, Ollama, LM Studio and vLLM.

FOR CLOSED CODING-MODEL VENDORS Recursive itself says Flint “may need more guidance than the larger closed-source coding models you’re used to.”

FOR AI RESEARCH LABS Recursive says its Eureka Machine “contributed throughout Flint’s development,” from architecture search to hardware-specific code.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework argues that the winning strategy in every major platform shift has followed one pattern: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, Recursive opens Flint’s weights under Apache 2.0, while the Eureka Machine that built them is described only in Recursive’s own posts. The framework was published in March 2026 and does not discuss Recursive or Flint.

What Is Not Established

The benchmark scores and speeds are Recursive’s own. The model card says evaluation setups vary across sources, and we found no independent test of Flint in what we read.

Recursive does not say how much of the work the Eureka Machine did versus its human team, beyond describing contributions across the stack. The model card documents a 262,144-token context configuration, while the launch post says Flint supports up to 512k tokens and recommends staying within 256k.

We read Techmeme’s summary of WIRED, not WIRED’s article itself. TechCrunch’s May report gave the $650 million figure but, in what we read, not a valuation.

The Bottom Line

Recursive, the Richard Socher-led lab that TechCrunch reported raised $650 million in funding in May, released Flint on 21 October 2026 as its first model. It is an open-weight coding model that runs locally, and the company says a small team and its Eureka Machine built it from scratch in under four months.