Recursive launches Flint, a code model created by a research AI

The first public model from Recursive, Flint, arrives with speed metrics on Apple Silicon and Nvidia, and is available for free under Apache 2.0. The claimed performance has not yet been independently verified, while the startup aims for self-improvement driven by its Eureka Machine system.

Performance to Validate and Hypothetical Self-Improvement

The announced results for Flint have not yet undergone independent verification. Flint represents a first concrete outcome, but it does not prove a complete recursive self-improvement. It remains to be seen whether the approach will yield self-accelerated improvement, and the expected demonstration would be that a model derived from the Eureka Machine improves the Eureka Machine itself. The actual scope of AI-driven research automation remains debated, with a study on over 700 task journals indicating that while agents handle large portions of the work, critical decisions remain predominantly human. Oriol Vinyals considers assisted self-improvement realistic, while deeming an explosion of intelligence unlikely. Recursive also acknowledges that Flint requires more guidance on complex tasks than proprietary models, and recommends breaking down significant tasks into smaller steps.

MoE Architecture and Measured Speeds on Apple and Nvidia

Flint is based on a mixture-of-experts architecture comprising 40 billion parameters, with 2.5 billion activated per token. On a machine equipped with an M4 Pro with 48 GB, it achieves the same results as significantly slower dense models ranging from 27 to 31 billion parameters. It is designed to compete with dense models of 25 to 40 billion while running more than five times faster. On a MacBook Pro with M4 Pro, it exceeds 125 tokens per second in short contexts, and surpasses 150 tokens per second on an Nvidia RTX Spark with 64 GB. Compared to similar models, inference is claimed to be up to twice as fast. The model handles contexts up to 512,000 tokens, with a recommendation of 256,000. Recursive also claims peak performance on coding benchmarks among MoE models with fewer than 50 billion parameters, results that remain to be verified.

What the Eureka Machine Decided and Built for Flint

According to Recursive, the Eureka Machine autonomously chose the architecture, quantization, and computation kernels, automating the entire research cycle by exploring multiple approaches in parallel. For Flint, it contributed to training data, optimizers, architecture design, and inference tuning. After evaluating different expert configurations from an initial version with 35 billion parameters, of which 3 billion were active, the system selected an option totaling 40 billion parameters and 2.5 billion active, deemed the best balance between speed and performance. For quantization, the output layer was reduced to four bits accompanied by additional distillation; Recursive indicates that about half the number of training steps was needed compared to the previous method, with a gap of approximately 0.7 points compared to the 16-bit version on benchmarks. The Eureka Machine also generated optimized kernels for Apple Silicon and Nvidia RTX Spark; on M4 and M5, a megakernel Metal merges the entire forward pass of generation into a single operation, with these optimizations aimed at maintaining inference speed.

Local Execution, Compatible Engines, and Open Distribution

Flint works with execution environments like llama.cpp, Ollama, LM Studio, vLLM, and SGLang. Recursive also publishes rec, an inference engine dedicated to Apple Silicon. The model can run locally without sending code to external servers and has been optimized for Apple and Nvidia hardware. The weights are offered for free under the Apache 2.0 license on Hugging Face.

Record Funding and Stated Goal Towards Superintelligence

Founded by Richard Socher along with Tim Rocktäschel and others, Recursive emerged from the shadows in May 2026 by raising $650 million for a valuation of $4.65 billion, with no concrete technical results to show at that time. Flint is the first public model resulting from its automated research process. The company aims for systems capable of supporting their own development, with the stated ambition of a superintelligence that continuously improves. Each new model is expected to enhance the research capabilities of the Eureka Machine and accelerate the next generation, with projected expansion into other scientific fields. At the same time, other labs are already using AI to design models: Anthropic claims that Claude handles a significant portion of its own development, and OpenAI indicates deploying agents throughout its research pipeline. The steps that Recursive wishes to automate fit within this movement.