Recursive Releases Flint, a 40B Open-Weight Local Coding Model for Agents

A New Local Coding Model

Recursive Superintelligence has released its first AI model, called Flint. It is a large language model designed specifically for coding agents — programs that investigate codebases, fix bugs, write tests and add features on their own. The model is open-weight and licensed under Apache 2.0, meaning anyone can download, study and modify it ¹.

Flint is a mixture-of-experts (MoE) model. This means it has many smaller "expert" networks inside it, but only a small portion processes each piece of text at a time. It has 39.8 billion parameters in total, but only about 2.5 billion are active for each token it generates. That makes it much faster and lighter to run than models where all parameters are always used ¹.

Built by an Automated Research System

According to Recursive, the Eureka Machine—a system that automates research—was responsible for constructing almost every component of the model. The system proposes ideas, implements them, tests them and then decides what to try next. Among its contributions was a creative approach to compression: it quantized (compressed) the output layer of the model to just 4 bits, which is unusual. Normally this part stays uncompressed. An offline training step using the original full-precision model as a teacher recovered most of the lost accuracy, and it nearly halved training time compared with standard methods ¹.

The team also wrote custom code for processing low-bit math, MoE operations, attention and prompt handling. On Apple M4 and M5 chips, a special kernel combines the entire decoding step into one operation, making it very fast ¹.

Performance and Benchmarks

Flint ranks at the top of Recursive's chart for MoE models under 50 billion parameters on agentic coding tasks, but it trails the strongest dense models — models that use all their parameters at once ¹. On SWE-bench Verified, a benchmark measuring how well models can solve real software engineering problems, Flint scored 74.2%. A dense model from Qwen, Qwen 3.8, scored 73.0% on Terminal-Bench, a different coding challenge, while Flint scored 54.9% . On LiveCodeBench, Flint achieved 80.3%, compared with 80.4% for Qwen 3.5 MoE .

Recursive states that the model runs at speeds exceeding 125 tokens per second on an M4 Pro MacBook Pro and surpasses 150 tokens per second on an NVIDIA RTX Spark, which the company says is up to double the performance of similar models ¹. It also claims Flint runs more than five times faster than dense 25-to-40-billion-parameter models with similar coding scores .

Hardware Requirements

The 4-bit version of Flint takes up about 23 gigabytes and runs on Apple Silicon Macs with at least 48 GB of memory — such as the M4 Pro, M5 Pro or M5 Max — or on NVIDIA GPUs with 48 GB of memory ¹. The full-precision version requires an 80 GB checkpoint and needs a server-grade GPU like the NVIDIA B300 or H200 . Flint is also compatible with popular local model tools including Llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang .

Flipped: The World's First Open Source, Open Revenue AI Model Maker

Key facts

• Flint has 39.8B total parameters with about 2.5B active per token
• Runs on 48GB Apple Silicon or NVIDIA GPUs, with 4-bit weights taking ~23GB
• Scores 74.2% on SWE-bench Verified and 95.1% on IFEval
• Built using Recursive's automated Eureka Machine research system
• Open-weight under Apache 2.0 license

Sources

1. Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params — marktechpost.com