Recursive Open-Sources Flint-40B-A2B: AI-Designed MoE Model for Local Coding Agents

Recursive Open-Sources Flint-40B-A2B: A Locally Optimized MoE Coding Agent

Recursive, the team led by Tian Yuandong, has released Flint-40B-A2B, a mixture-of-experts (MoE) model designed for local coding agents. The model was developed end-to-end — including tokenizer, pre-training, mid-training, supervised fine-tuning, and reinforcement learning — in under four months using Recursive's proprietary AI automatic research system, Eureka Machine.

Performance on Local Hardware

Flint is optimized for on-device inference. On a MacBook Pro, it generates over 125 tokens per second; on NVIDIA RTX Spark devices, it exceeds 150 tokens per second — up to 2× faster than comparable models.

Autonomous Research Capabilities Demonstrated

Beyond code generation, Flint can execute a full autonomous research loop on a laptop. Given a task and measurable evaluation metric, it proposes hypotheses, writes code, runs experiments, and iteratively improves solutions based on feedback.

Circle Packing Optimization

Without visual input, Flint arranges circles within a square to optimize layout. It adjusts positions and radii via code, tests different initialization strategies, and continues optimizing based on scores.

Factory Scheduling

Flint reorders job execution sequences to minimize machine idle time. Compared to a round-robin baseline, overall completion time improved by 26%.

Bike-Sharing Demand Prediction

Using the public UCI Bike Sharing dataset, Flint predicts rental demand from weather and date features. It explored log and quadratic time features combined with residual tree models, reducing validation RMSLE by 30.8% and hold-out test RMSLE by 9.1%.

Eureka Machine's Role in Model Development

Flint itself is a product of AI-driven research. Eureka Machine participated across the development pipeline:

Architecture search: Starting from a 35B total / 3B active parameter baseline, the system explored expert configurations and converged on 40B total / 2.5B active parameters.

Quantization: The output layer was compressed to 4-bit, with additional distillation to recover accuracy.

Inference kernels: Hardware-specific compute kernels were generated for Apple Silicon and NVIDIA GPUs to maximize throughput.

Recursive Self-Improvement (RSI) Progress

This work aligns with Recursive's long-term RSI vision: AI improves the AI development process, and stronger models then enhance the next round of automated research. However, the team emphasizes that Flint is a human–AI collaborative outcome, not yet fully autonomous recursive self-improvement.

Flint demonstrates that an AI research system can participate in model training, and the resulting model can itself perform autonomous algorithm exploration on local hardware — showing two linked stages of AI-developing-AI.