Recursive Introduces Flint, a Mixture-of-Experts (MoE) Model Built for Fast Agentic Coding on Local Hardware

Recursive, building the Eureka Machine, the self-improving superintelligence that automates the scientific method, led by Richard Socher and the team, has introduced Flint, a mixture-of-experts (MoE) model built for fast agentic coding on local hardware. Flint has 40 billion total parameters and activates just 2.5 billion per token, built from scratch in under four months.

Recursive’s Flint, an MoE model with fewer than 50 billion parameters, matches dense models’ coding performance while running over 5x faster. Its optimized inference engine delivers 125+ tokens per second on an M4 Pro MacBook Pro. Flint was developed using Recursive’s Eureka Machine, an automated system for AI research and development.

State-of-the-Art Local Agentic Coding Performance.

Flint delivers state-of-the-art local agentic coding performance among MoE models under 50 billion parameters, rivaling similarly sized dense models while running significantly faster.

Designed for private, local coding, Flint supports context windows up to 512k tokens, with 256k recommended for optimal results. Its inference stack is optimized for faster execution on local hardware, while complex tasks benefit from explicit instructions and smaller subtasks.

Fast Inference on Your Local AI Device.

Flint delivers fast, capable local AI coding on Apple Silicon and NVIDIA RTX Spark, enabling offline work, data privacy, and zero API fees. Its optimized inference engine runs up to 2x faster than comparable models and over 5x faster than similarly capable dense models.

How Flint Was Built With Automated Research?

Recursive’s Eureka Machine automates AI research by generating ideas, running experiments, evaluating results, and refining future approaches in parallel.

The system supported Flint’s development across data, training, optimization, infrastructure, and inference. Its discoveries enabled a small team to build the model from scratch and achieve state-of-the-art benchmark results in under four months.

Flint’s local inference speed demonstrates how automated research can drive improvements across the entire AI model-building stack.

Architecture: Recursive’s Eureka Machine explored expert configurations to optimize local inference, resulting in Flint’s 40-billion-parameter MoE architecture with just 2.5 billion active parameters per token.

Quantization: Automated experiments produced an NVFP4 model that nearly halves quantization training steps while scoring within 0.7 points of the 16-bit model across benchmarks.

Kernels: Eureka Machine developed hardware-specific optimizations for Apple Silicon and NVIDIA RTX Spark, including a Metal megakernel that combines the entire decoding pass on M4 and M5 chips.

Together, these advances demonstrate how automated research can compound improvements across architecture, training, and inference to deliver faster, more efficient local AI.

Running Flint.

Flint’s weights are available on Hugging Face under the Apache 2.0 license for local use and customization. The model supports frameworks including llama.cpp, Ollama, LM Studio, vLLM, and SGLang. Recursive has also released rec, a lightweight inference engine optimized for Flint on M4 Pro and M5 Pro/Max hardware.

What’s Next?

Recursive aims to build self-improving AI systems that automate knowledge discovery, starting with model development. Flint marks the first step, built from scratch in under four months with Eureka Machine automating research and experimentation across the model-building stack.

The next phase focuses on using more capable models to improve Eureka Machine itself, accelerating development while reducing costs, alongside advancing local AI models and the infrastructure needed to run them.