AI startup Recursive has unveiled Flint, an open-weights mixture-of-experts (MoE) model tailored for local agentic coding. The model contains 40 billion total parameters with 2.5 billion active per token, designed to rival dense models in the 25B to 40B range while executing up to five times faster. Flint was developed over four months using Recursive’s proprietary “Eureka Machine,” an automated system that autonomously proposes, executes, evaluates, and iterates on research experiments across architectures, optimizers, and quantization schemes.
Recursive claims Flint leads benchmark tests among sub-50B parameter MoE models on coding evaluations like SWE-Bench Pro v2 and Terminal-Bench 2.1, though the results await independent verification. On local hardware, the company reports token generation speeds exceeding 125 tokens per second on Apple M4 Pro chips and 150 tokens per second on NVIDIA RTX Spark hardware. The optimizations were driven by the Eureka Machine, which implemented specialized kernels, including a custom Metal megakernel for Apple Silicon.
The release represents the first public technical deliverable from Recursive, which emerged from stealth in May 2026 after raising $650 million at a $4.65 billion valuation. Flint’s model weights have been made publicly available under an Apache 2.0 license, supporting context windows up to 512,000 tokens across runtimes like llama.cpp, vLLM, and Ollama.
Why it matters
Serves as an initial proof-of-concept for automated AI research platforms capable of designing, training, and optimizing models autonomously.
Provides developers a high-speed, open-source 40B MoE model optimized specifically for running localized agentic coding workflows.
Demonstrates advanced local hardware acceleration techniques that significantly improve token generation speeds on consumer devices.
Source: the-decoder.com



