Recursive releases Flint, a coding model it says a laptop can run
The company, Wednesday: about 39.8 billion parameters, with roughly 2.5 billion active per token, under Apache 2.0 on Hugging Face. Richard Socher told WIRED that nearly every part of the pipeline was done by an automated research system. TechCrunch put Recursive’s own speed test at about 131 tokens a second on an M4 Pro. The launch post says the context window goes to 512,000 tokens. The model card lists 262,144.
Feature4 minIn-depth
James Okonkwo4 min read
- Recursive
- Flint
- Coding agents
- Open weights

Photo: WIRED Staff / Getty Images, via WIREDSource page
Recursive released Flint on Wednesday, an open-weight coding model the company says is small enough to run on a laptop. The Hugging Face card puts it at 39.8 billion parameters, with about 2.5 billion active per token once the output layer is counted, and about 2.2 billion in the rest of the network that runs on each token. Recursive’s launch post rounds that to 40 billion total and 2.5 billion active. The weights are under the Apache 2.0 license. Aditya Mehta at TechCrunch, in a story timestamped 1:27 p.m. Pacific, called Flint the company’s first model and wrote that it is free to download.
The claim about how it was built is the one Richard Socher, Recursive’s chief executive and a cofounder, gave Will Knight at WIRED. “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” he said. The system is called the Eureka Machine. Recursive’s post says it proposes ideas, implements them, tests them and uses the results to choose the next try, and that it worked on data, training recipes, infrastructure and inference. The company says a small team built Flint from scratch in under four months. Socher told Knight the system that produced Flint cannot get out from under human control, and that the longer aim is for the Eureka Machine itself to improve without a person writing each step, with people still able to stop it. “We hope that this shows people that indeed we have all the steps in place,” he said. WIRED reported a $650 million raise at a $4.65 billion valuation, from Google Ventures, Greycroft, Nvidia and AMD Ventures. TechCrunch has the same round, from the company’s emergence from stealth in May, and names GV, Nvidia and AMD Ventures.
The company’s scores, and two context numbers
The model card’s release scores, which Recursive marks as its own measurements, are 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2 and 74.2 percent on SWE-bench Verified. Sampling on the card is temperature 1.0, top-p 0.95 and top-k 20. Terminal-Bench uses a harness called Terminus 2. The SWE-bench runs use mini-swe-agent. The agent scores are an average of at least five attempts. The card says a Terminal-Bench solution found by searching the web counts as a failure, and that scores reported by other labs may not use that rule. Other numbers on the card, used for comparison, are what those developers published. Recursive’s post says Flint leads mixture-of-experts models under 50 billion parameters on the three tests, and that it matches dense models of 25 to 40 billion parameters while running more than five times as fast. Mehta, reading an average Recursive published across six tests, wrote that only Alibaba’s Qwen3.8 27B scored higher. Those ranks are the company’s.
Speed, in TechCrunch’s account of Recursive’s tests, was about 131 tokens a second on a MacBook Pro with an M4 Pro chip, against roughly 58 to 69 for Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same machine. Recursive’s post says more than 125 tokens a second on an M4 Pro at short contexts, and more than 150 on a 64-gigabyte Nvidia RTX Spark. Knight tried it on a Mac with 48 gigabytes of memory. He wrote that it finished the small programs he asked for, including a command-line to-do list, and that models of this size are often slow or wrong next to Claude or Codex in his own use. That is one reporter’s afternoon. It is not the table.
The context window is where Recursive’s own pages do not match. The launch post says Flint supports up to 512,000 tokens and recommends staying inside 256,000 for coding. The model card’s documented configuration is 262,144 tokens, prompt and output together. Someone who downloads the checkpoint is looking at the card. The post’s 512,000 is not the number that card lists.
The software around the weights is spelled out in both places. Recursive says Flint works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, and that it is also shipping rec, an inference engine for M4 Pro and for M5 Pro and Max machines. The card says the 16-bit checkpoint is about 80 gigabytes and, with the SGLang kit in the repository, runs on one B300 or one H200. Code for Hugging Face Transformers is not in the repo yet. On compression, the launch post says the Eureka Machine quantized the output layer to 4 bits, which the company calls a break with common practice, then trained the compressed model on its own answers with the 16-bit model as the teacher. The resulting NVFP4 model, Recursive says, lands within 0.7 points of the 16-bit model on its benchmarks, and the extra stage nearly halved the number of training steps against the company’s quantization-aware baseline. TechCrunch put a different figure on the same work: low-bit training shrank Flint by 71 percent with almost no loss of quality. The 71 percent is TechCrunch’s summary of Recursive.
The card’s warning is plain. Fine-tuning included safety examples, and the model can still write faulty code or call a tool it should not. A strict format for those calls does not stop the choice. Recursive tells users to split hard jobs and to expect to give more direction than they would to a larger closed coding assistant. Mehta added a size comparison that does not need a benchmark: Meta’s Muse Spark Open, which he wrote was released last week, uses 816 billion parameters and needs at least a terabyte of memory. Flint is the smaller release, with the weights public. Socher’s account of the research loop still has a person who can turn it off.


