Recursive trains a code model with an AI that participates in its own design

Recursive releases Flint, a 40-billion-parameter AI programming model designed with Eureka Machine to run fast locally.

An artificial intelligence model developed with the help of another system designed to improve training methods, architectures, and computational performance. That is the approach behind Flint, the first coding model from Recursive, the research company founded by Richard Socher. Built from scratch in less than four months by a small team of researchers and AI agents, Flint contains nearly 40 billion parameters while activating approximately 2.5 billion for each token. Recursive reports generation speeds above 125 tokens per second on an M4 Pro MacBook Pro and coding performance competitive with substantially more computationally expensive models. Its weights are available under Apache 2.0, alongside an inference engine optimized for personal computers. A Coding Model Designed to Stay on Your Computer

Advanced AI coding assistants generally rely on models hosted in data centers. They can examine repositories, fix bugs, and use developer tools, but often require an internet connection and involve sending code to remote infrastructure. Local execution offers an alternative for developers working with confidential projects or seeking to avoid usage-based API charges, provided the model is capable and efficient enough.

Flint was designed around that trade-off. It supports repository investigation, bug fixing, feature implementation, and test creation, including tool use across multiple conversational turns. Its purpose is not merely code completion, but participation in engineering tasks that require several steps.

The architecture uses a mixture-of-experts (MoE) design. Unlike a dense model, which activates its entire network for each token, Flint selects a subset of specialized components. It has 39.8 billion total parameters, with approximately 2.5 billion active per token, including its output head. This reduces the computation required during generation, although the complete set of model weights must still be accessible in memory.

Flint consists of 49 layers, including two dense layers and 47 MoE layers. Each expert layer contains 256 routed experts, selecting six per token alongside a shared expert. The architecture also combines local and full-attention mechanisms to manage longer contexts more efficiently. More Than 150 Tokens per Second on Local Hardware

Recursive optimized the inference pipeline from weight representation to hardware-specific computation. Using its optimized engine, the company reports 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, measured at short context lengths.

According to Recursive's benchmarks, Flint can run up to twice as fast as selected models in its class. The company also claims speeds more than five times higher than dense models with 25–40 billion parameters and comparable coding benchmark results. Those comparisons depend on the specific models, quantization methods, inference engines, and hardware configurations involved.

Recursive has released rec, an inference engine optimized for Flint on Apple M4 Pro and M5 Pro or Max systems. The Apache-licensed project complements support through llama.cpp, Ollama, LM Studio, Unsloth, vLLM, and SGLang.

Local inference can keep source code on the user's machine and eliminate per-request cloud inference fees. It still requires suitable hardware. The complete BF16 checkpoint occupies approximately 80 GB, while quantized variants substantially reduce memory requirements. Available capabilities also depend on the execution engine and its implementation of tool calling. Strong Results on Agentic Coding Benchmarks

Recursive evaluated Flint across several benchmarks measuring programming, problem-solving, and tool use. Its Hugging Face model card reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.

These evaluations measure different abilities. Terminal-Bench focuses on tasks carried out through terminal environments, while SWE-bench variants test the resolution of software engineering problems involving code repositories. Recursive also evaluates Flint on LiveCodeBench v6, IFBench, and IFEval to assess coding and instruction-following capabilities.

The company positions Flint as the strongest MoE model with fewer than 50 billion total parameters on its agentic coding comparisons. It also reports results competitive with larger dense models, while emphasizing substantially higher inference speeds.

Those figures require qualification. Flint's published scores come from Recursive's own evaluations and have not yet been comprehensively reproduced independently. Benchmark settings are not necessarily identical to those used for results published by competing developers. On Terminal-Bench, for example, solutions obtained through web retrieval are counted as unsuccessful, which can affect direct comparisons.

Agentic evaluations involve at least five attempts and timeouts of up to five hours. They therefore measure whether a system can complete a task within a defined environment rather than the immediate quality of a single answer. Recursive also acknowledges that Flint may require more explicit instructions and smaller subtasks than larger proprietary coding models. Eureka Machine Automates Parts of AI Research

The most distinctive aspect of Flint's release is how the model was developed. Recursive credits Eureka Machine, its automated research infrastructure, with contributing to much of the development process. The system proposes hypotheses, implements experiments, runs evaluations, and uses the results to decide which directions to investigate next.

Unlike a coding agent implementing one requested feature, Eureka Machine addresses problems in the design of AI models themselves. It can explore multiple possibilities in parallel, compare outcomes, and combine promising discoveries. Human researchers still define objectives and supervise development, but agents perform part of the experimentation.

For Flint, Recursive