AI startup Recursive has released Flint, its first language model, and says an internal AI system called Eureka Machine took part in nearly every step of building it. According to The Decoder, the system proposed ideas on its own, built them, tested them and planned follow-up experiments based on the results. Flint is a mixture-of-experts coding model designed to run quickly on local Apple and Nvidia hardware, it was built in four months, and it is available on Hugging Face under the Apache 2.0 license. The benchmark claims come from Recursive and have not been independently verified.

What Flint is

Flint is a mixture-of-experts (MoE) model with 40 billion total parameters, of which only 2.5 billion are active for each token. In an MoE design, a routing step sends each token to a small subset of specialist sub-networks, so the model can hold a lot of knowledge while spending the compute of a much smaller one. Recursive built Flint for agentic coding, meaning tasks where a model plans, edits files and runs commands, on local machines instead of in a data center.

The company says Flint is meant to compete with dense models in the 25 to 40 billion parameter range while running more than five times faster. Reported speeds:

  • More than 125 tokens per second at short contexts on a MacBook Pro with an M4 Pro chip.
  • More than 150 tokens per second on an Nvidia RTX Spark with 64 GB of memory.
  • Up to twice the inference speed of similar MoE models, according to Recursive’s charts.

Recursive also says that on an M4 Pro with 48 GB of memory Flint matches the benchmark scores of much slower dense models in the 27 to 31 billion parameter range.

The benchmark claims

Recursive says Flint leads coding benchmarks including SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified among MoE models under 50 billion parameters, naming Qwen3.6-35B-A3B and Laguna S 2.1 as comparison models. Across six coding and instruction-following benchmarks it leads similar sparse MoE models, though The Decoder notes that dense models such as Qwen 3.8 score higher in some tests. These are the company’s own results. Readers should wait for third-party evaluations before treating the rankings as settled, especially for agentic coding benchmarks, which are sensitive to scaffolding and settings.

The Eureka Machine

The more unusual part of the announcement is the process. Recursive says Eureka Machine selected the architecture, the quantization scheme and the compute kernels on its own. The company’s stated long-term goal is to create AI systems that can take charge of their own further development, a theme that sits at the center of debates about recursive self-improvement and the path to more general AI.

It is important to be precise about what has been shown. Automated research, where a system proposes and runs experiments, is an active area, and Flint is presented as the first public result of Recursive’s version. A single model built in four months does not demonstrate open-ended self-improvement. What it does show, if the account holds, is that an automated loop can contribute to producing a competitive small model. We do not yet have details on how much human supervision and compute the loop needed, or how many experiments failed, and the claims about the process come from the company.

Why it matters

Two trends meet in this release. The first is the rise of sparse, efficient models that run on laptops, which lowers the cost of private, offline coding assistance. The second is the automation of AI research itself. If a small team can use an automated system to produce a competitive model in months, the pace of model development could depend less on headcount and more on compute and evaluation quality. That cuts both ways for safety discussions: faster iteration helps builders, while less human oversight of the process makes it harder to audit what changed and why.

For developers, the practical points are simpler. The Apache 2.0 license allows commercial use, the model is designed for consumer-class hardware, and anyone can test it against their own repositories rather than rely on published scores.

What to watch next

  • Independent benchmark runs and reproductions of Recursive’s SWE-Bench and Terminal-Bench results.
  • A technical report describing what Eureka Machine did and how much human review was involved.
  • Community testing of Flint’s real-world agentic coding quality on local hardware.
  • Follow-up models from Recursive, and whether the automated process scales to larger systems.
  • How other labs describe and disclose the role of automated research in their own releases.

Key takeaways

  • Recursive’s Flint is a 40B-parameter mixture-of-experts coding model with 2.5B active parameters, released under Apache 2.0 on Hugging Face.
  • The company says its Eureka Machine system proposed, built and tested most experiments and chose the architecture, quantization and kernels.
  • Reported speeds exceed 125 tokens per second on an M4 Pro MacBook and 150 on an RTX Spark.
  • Benchmark leadership among MoE models under 50B parameters is claimed by Recursive and not yet independently verified.
  • One four-month model does not prove open-ended self-improvement, though it is a notable data point.

Frequently asked questions

What is a mixture-of-experts model?

It is a model that routes each token to a few specialist sub-networks, so only a fraction of the parameters is active at once. Flint uses 2.5 billion of its 40 billion parameters per token.

Can I run Flint on my own computer?

Recursive says it is optimized for fast local execution on Apple and Nvidia hardware. The model is on Hugging Face under the Apache 2.0 license.

Is Flint built entirely by AI?

No. Recursive says Eureka Machine was involved in nearly every step, but the company, and the humans in it, still directed the work. The extent of human oversight has not been detailed.