ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Recursive Releases Flint: A 40B Open-Source Coding Model Developed Primarily by Its Internal AI System

Recursive launches Flint, a 40B parameter MoE coding model developed largely by its internal AI system Eureka Machine, claiming leading performance among similar models with significantly faster inference.

2026-10-22 22:40 GMT+8

AI startup Recursive released Flint on October 22, 2026, its first language model. Flint is a mixture-of-experts (MoE) model designed for agentic coding tasks on local hardware, featuring 40 billion total parameters with 2.5 billion active per token. According to The Decoder, the development of Flint was driven almost entirely by Recursive's internal AI research system, "Eureka Machine," which autonomously proposed experiments, ran tests, and planned follow-up steps based on results, completing the four-month development cycle.

Recursive claims that Flint leads coding benchmarks such as SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified among MoE models with fewer than 50 billion parameters. These results have not yet been independently verified. In terms of performance, the company states that Flint aims to compete with dense models in the 25 to 40 billion parameter range while running more than five times faster. For instance, on an M4 Pro chip with 48 GB of memory, Flint generates over 125 tokens per second at short contexts; on an Nvidia RTX Spark with 64 GB of memory, it exceeds 150 tokens per second.

The Eureka Machine played a key role in architecture design, quantization strategy, and compute kernel optimization. It autonomously tested different expert configurations, ultimately settling on the 40B total / 2.5B active parameter design as the best tradeoff between speed and coding performance. Regarding quantization, the system broke common practice by reducing the output layer to four bits, recovering accuracy through an added distillation step so that the final NVFP4 model sits only 0.7 points behind the 16-bit version across benchmarks. Additionally, specialized compute kernels were built for Apple Silicon and Nvidia hardware, such as combining the entire forward pass into a single Metal megakernel operation on M4/M5 chips.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Flint supports context windows of up to 512,000 tokens (with a recommended cap of 256,000), and the model weights are available on Hugging Face under the Apache 2.0 license. Supported runtimes include llama.cpp, Ollama, LM Studio, vLLM, and SGLang. Recursive also released rec, its own inference engine for Apple Silicon. Despite its strong local execution capabilities, Recursive acknowledges that Flint requires more guidance on complex tasks than large proprietary coding models, recommending that users break bigger tasks into smaller steps.

Recursive was founded by Richard Socher and others including former DeepMind researcher Tim Rocktäschel. The company emerged from stealth in May 2026, raising $650 million at a $4.65 billion valuation. Flint is the first concrete result from its automated research process, with the long-term goal of using this approach to build superintelligence that keeps improving itself. However, Flint's success does not yet prove full recursive self-improvement; achieving that would require demonstrating that a model built with the Eureka Machine actually makes the Eureka Machine itself a better researcher.

Sources

  • the-decoder.comThe Decoder↗
Continue researching →