Flint AI: A Coding Model Built Largely by AI

Recursive, the San Francisco AI startup that came out of stealth in May 2026 with one of the largest seed-stage raises in the industry, released its first product on Wednesday. The model is called Flint. It is a free coding model designed to run entirely on a laptop, and Recursive says the bulk of the engineering behind it was carried out by its own automated research system rather than by human researchers.

The release matters for two separate reasons. For developers, Flint is a fast local coding model with open weights under an Apache 2.0 licence, which means it can be downloaded, run offline, modified and used commercially without API fees. For the wider industry, it is the first public output of a company whose stated mission is recursive self-improvement, the idea that AI systems can be used to build better AI systems, which can then build better ones still.

What Flint Is

Flint is a mixture-of-experts model with 40 billion total parameters. Only 2.5 billion of those parameters are active for any given token, which is what allows a model of this size to run on consumer hardware. Recursive says a small team of people and AI agents built it from scratch in under four months.

The model is aimed squarely at agentic coding, where the model takes multiple steps on its own rather than only completing a line of code. Recursive says Flint can be used for agentic workloads on machines with 48GB of RAM, and that it supports context windows of up to 512,000 tokens, though it recommends staying within 256,000 tokens for the best coding results.

Speed Claims: Roughly Twice as Fast as Comparable Models

Speed is the headline claim. In Recursive's own tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. The company says comparable models on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, managed between 58 and 69 tokens per second. On a 64GB Nvidia RTX Spark, Flint passed 150 tokens per second at short contexts.

Recursive also claims Flint runs more than five times faster than dense models in the 25 to 40 billion parameter range that reach similar scores on coding benchmarks. The company's published speed figures are measured as average tokens per second across a 128k-token session, with prefill excluded.

Those numbers come with a caveat. They were produced using Recursive's own optimised inference engine, which it is releasing as an open-source tool called 'rec', tuned for M4 Pro and M5 Pro and Max chips. Performance in third-party runtimes such as Ollama or LM Studio may differ.

Benchmark Results

Recursive says Flint is the best-performing mixture-of-experts model with fewer than 50 billion total parameters on three agentic coding benchmarks: SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. SWE-Bench Verified in particular measures how well a model can fix real bugs in real open-source repositories.

On a six-benchmark average that also includes LiveCodeBench v6, IFBench and IFEval, Recursive reports that only one model scored higher than Flint: Alibaba's Qwen3.8 27B, a dense model. Recursive's position is that Flint trades a small amount of raw accuracy for a large speed advantage over models of that type.

The evaluation setup Recursive published is specific. Agentic tests used agent timeouts of up to five hours and were averaged over at least five attempts. Terminal-Bench 2.1 was run with internet access, but any solution obtained through retrieval was counted as a failure. SWE-Bench Pro v2 and SWE-Bench Verified were run with mini-swe-agent and no internet access. Sampling used a temperature of 1.0, top-p of 0.95 and top-k of 20.

How the Eureka Machine Built It

Recursive's automated research system, which it calls the Eureka Machine, proposes ideas, implements them, runs experiments, and uses the results to decide what to try next. It works on several directions in parallel and combines ideas that show promise. According to the company, the system contributed to nearly every layer of Flint's development, from data and optimisers to training recipes, infrastructure and inference.

Recursive has described three specific contributions in detail.

1. Architecture search. Starting from a baseline of 35 billion total parameters and 3 billion active, the Eureka Machine explored different expert configurations and widths. It settled on 40 billion total and 2.5 billion active as the best projected balance of speed and coding performance.
2. Quantisation. After training, the system searched for recipes to shrink the model's memory footprint. Against common practice, it quantised the output layer to 4 bits, then recovered the lost accuracy with an offline distillation stage in which the quantised model trained on its own pre-sampled responses with the 16-bit model as teacher. Recursive says this nearly halved the number of training steps compared with its quantisation-aware training baseline, and the final NVFP4 model scores within 0.7 points of the full-precision version. Overall, low-bit training shrank Flint by 71 percent.
3. Custom kernels. The system wrote specialised kernels for low-bit projections, mixture-of-experts operations, attention and prompt processing on Apple Silicon and the RTX Spark. On M4 and M5 chips it produced a Metal megakernel that fuses the entire decode forward pass into a single operation.

Recursive has not said how many researcher-hours the system saved. It is not the only lab leaning on AI for this kind of work. OpenAI said in September that by mid-August its AI agents were logging 3.1 workdays for every single day worked by a human researcher.

How It Compares With Meta's Muse Spark Open

Flint arrived one week after Meta released Muse Spark Open, an open-weight model built to run on Macs. The contrast in scale is stark. Muse Spark Open uses 816 billion parameters and needs at least a terabyte of memory. Flint, at 40 billion total parameters and 2.5 billion active, is designed for ordinary laptops with far less memory. Recursive's speed chart notes that Muse was measured at 124k tokens rather than the full 128k session used for other models.

Early Hands-On Results

Wired's Will Knight tested Flint on a Mac with 48GB of memory, using it to build several small programs including a to-do list app and a tool that searches the web for research material. Knight reported that smaller coding models he had tried previously often ran into errors or were slow compared with Claude or Codex, while Flint completed every coding request he gave it.

Recursive itself has been candid about the model's limits. It says Flint may need more guidance than the large closed-source coding assistants developers are used to, and it recommends breaking complex work into smaller subtasks with explicit instructions.

Availability

Flint's weights are available now on Hugging Face under the Apache 2.0 licence, which permits free use, modification and redistribution, including for commercial purposes. The model works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. Recursive is also shipping its own lightweight inference engine, rec, on GitHub, along with setup documentation.

Because the model runs locally, code never leaves the developer's machine. That removes two costs associated with cloud coding assistants: the per-token usage fees that have become a growing line item for engineering teams, and the privacy exposure of sending proprietary code to a third-party server.

The Company Behind It

Recursive, formally Recursive Superintelligence, Inc., was founded in 2025 and emerged from stealth on May 13, 2026 with $650 million in funding at a $4.65 billion valuation. Backers include GV, Nvidia and AMD Ventures. The New York Times reported at the time that the founding team includes former Google, Meta and OpenAI researchers.

The company is led by CEO Richard Socher, who was previously chief scientist at Salesforce and is also the founder of the AI search engine You.com. Socher published a book in September 2026 titled The Eureka Machine, laying out his argument that self-improving AI could accelerate breakthroughs in science, medicine and economics.

In June 2026 Recursive signed a $410 million compute agreement with Amazon Web Services. At the time the company told TechCrunch the deal was 'less about headcount and more about agent count', a line that now reads as a preview of how Flint was built.

Recursive published its first technical results from the Eureka Machine on June 11, 2026, covering model training and GPU kernel benchmarks. Flint is the first time the system has been used end to end to ship a model.

Why Recursive Self-Improvement Is Contested

Recursive self-improvement is one of the most debated concepts in AI. Supporters see it as the fastest path to scientific progress. Critics worry that a system which improves itself could move beyond human understanding or control. Recursive's current implementation is deliberately narrower than the full version of the idea: the Eureka Machine operates as a research assistant supervised by people, and its output so far is a model small enough to run on a laptop.

Socher has said the longer-term goal is full recursive self-improvement with humans overseeing the loop and retaining the ability to stop it. The company's stated bet is that each new model makes the Eureka Machine a better researcher, which in turn lets it build the next model faster and deliver more capability for the same cost.

One point worth watching is independent verification. Recursive's benchmark figures and speed comparisons are self-reported, and the most interesting claim, that the research system itself is improving, has not yet been measured by a third party. A fast open-weight coding model is a useful product on its own. Evidence that an AI research system is outperforming its own prior generation on new problems would be something else entirely.

What Comes Next

Recursive says its roadmap includes increasingly capable local models and the infrastructure to run them, as part of a broader effort to automate AI research and development. If the approach works as intended, each generation should be able to take on longer stretches of work with less human involvement.

For developers, the immediate question is simpler: whether a free 40-billion-parameter model that fits on a laptop and runs at over 125 tokens per second is good enough to replace a paid cloud assistant for everyday work. The early signals from benchmarks and hands-on testing suggest it is at least close.