Flint, from Recursive: a free coding model that runs on a laptop

On October 21, 2026, the startup Recursive released Flint, its first model. It is a free programming assistant with open weights under the Apache 2.0 license, small enough to run on a laptop. For a mid-sized company with a development team, the combination has three practical effects: zero licensing cost, no per-use fees, and source code that never leaves the machine.

Before getting carried away, it's worth separating what is proven from what is promise. Almost all the numbers below come from Recursive itself. The material gathered includes no independent evaluation.

What Flint is

The full name is Recursive Flint. The company describes it as a "mixture-of-experts" model (several internal specialists, only a few activated at a time) built for fast agentic programming on local hardware. Agentic, here, means the model carries out tasks in sequence rather than just answering questions.

The size figures explain why it fits on an ordinary machine: 40 billion parameters in total, but only 2.5 billion active per token. In practice, it thinks with a fraction of its size, and that matters for speed.

Another efficiency figure, also reported by the company: low-precision training cut the model's size by 71%, with minimal loss of quality.

Speed: what the company claims

According to Recursive, in short contexts Flint exceeds 125 tokens per second on a MacBook Pro M4 Pro and 150 tokens per second on a 64GB NVIDIA RTX Spark. Press coverage cites about 131 tokens per second on the MacBook, attributed to the company's own tests. Recursive says it can be up to twice as fast as comparable models.

As for competitors, the only second-hand numerical comparison in the material comes from theoutpost.ai: 58 to 69 tokens per second for Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning on the same machine. Those numbers do not appear in the primary source. No quality comparison among the three has been published.

Three caveats weigh on the decision:

• The speeds apply to short contexts. Flint accepts up to 512 thousand tokens of context, but the company recommends staying within 256 thousand.
• The company also released "rec", a lightweight engine optimized for Flint on Apple M4 Pro and M5 Pro/Max chips. The published speeds depend on this optimized engine.
• The minimum memory required for the 40B model has not been disclosed. The tests cite 64GB machines in the case of the RTX Spark.

Benchmarks: no scores, no conclusion

Recursive says Flint leads models of the same type with fewer than 50 billion total parameters on agentic coding tests: SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified. It also says it rivals dense models of 25 to 40 billion parameters at coding while running more than five times faster.

The scores were not published in the material, and no outside party has confirmed them. For anyone deciding based on money and risk, this means the only benchmark that counts is the one run on your own code.

Largely made by AI

Recursive states that the model was built from scratch in under four months by a small team of people and agents. The automated research system, called Eureka Machine, took part in almost every stage, from pre-training to inference. CEO Richard Socher, a former chief scientist at Salesforce, said that almost every aspect of the process was carried out and improved by AI. The company calls the release a first public test of its research agenda in recursive self-improvement.

Human researchers supervised and picked the promising results. remio.ai notes that the origin is hard to audit: the download lets you examine the model, but not the full process or the experiments that failed. For a manager, that discussion is secondary. What matters is whether the model delivers on your code.

What changes in operations

The business case is simple. A local assistant works offline, the data stays on the machine, and there is no API fee. That holds only if the model actually runs locally. For a company hesitant to paste client code into an external service, the risk of leakage by that route disappears.

The model is available on Hugging Face and works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang. The license cost is zero. The real cost lies elsewhere: the time of whoever sets it up, the right machine, and validating the output.

How to test without betting big

The material includes no productivity-gain data, so measurement is up to each company. In my reading, the safest path has four steps:

1. Pick a single pain point. Automated tests, refactoring, or documentation are repetitive tasks that are easy to review.
2. Set aside one machine and two developers. One owner per pilot keeps the test from dying halfway.
3. Measure by sprint. Record the hours spent on the task before and during the pilot, and how many times the result had to be redone.
4. Decide with numbers. If hours came back to the team and rework stayed low, expand. If not, you spent a few weeks and no licenses.

Flint still has no independent evaluation and depends on specific hardware to reach the published speeds. Even so, the cost of testing is low, and the result can be measured in hours returned to the team.

Recursive is a startup founded in 2025. Secondary aggregators mention US$650 million raised at a US$4.65 billion valuation, but no primary source in the material confirms the figure.