AI startup Recursive has released Flint, a free coding model designed to run locally on laptops, marking its first public demonstration of AI-assisted model development. The company said its automated research system, Eureka Machine, helped a small team build Flint from scratch in less than four months by assisting with model design, training, and performance optimization.
The model launched on October 21, 2026, is available under the Apache 2.0 open-source license. Recursive, led by former Salesforce chief scientist Richard Socher, emerged from stealth in May after raising $650 million at a $4.65 billion valuation from investors including GV, Nvidia, and AMD Ventures.
Flint Runs Locally With 40 Billion Parameters
Flint uses a mixture-of-experts architecture containing 40 billion parameters, although only 2.5 billion are activated for each token it processes. This design reduces computing requirements while allowing the model to handle coding tasks on compatible laptops and workstations.
Because Flint operates locally, developers can work offline, keep their source code on their own devices, and avoid usage-based cloud API fees. The model is available for download on Hugging Face, alongside support for development tools including Ollama, LM Studio, llama.cpp, vLLM, and SGLang.
According to Recursive’s internal benchmarks, Flint generated approximately 131 tokens per second on a MacBook Pro equipped with Apple’s M4 Pro chip. Comparable models, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, reportedly achieved between 58 and 69 tokens per second under similar conditions.
Flint also exceeded 150 tokens per second on Nvidia’s RTX Spark hardware. Recursive said it outperformed other mixture-of-experts models with fewer than 50 billion parameters on SWE-Bench Verified, while Alibaba’s Qwen3.8 27B achieved a higher average across six evaluations.
Eureka Machine Automates AI Model Development
Recursive developed Flint using Eureka Machine, an AI research system that proposes ideas, conducts experiments, evaluates results, and determines which approaches to test next. The system contributed to model architecture, training procedures, data processing, and inference optimization.
According to Socher, Eureka Machine developed a more efficient model architecture and low-bit training techniques that reduced Flint’s memory requirements by 71% with minimal performance loss. It also generated specialized computing programs to improve execution speed on Apple Silicon and Nvidia hardware.
Recursive aims to use this approach for recursive self-improvement, where AI systems assist in developing increasingly capable models. However, Eureka Machine currently operates with human oversight rather than independently controlling the entire research process.
Recursive Plans Further AI Research Automation
Recursive acknowledged that Flint may require additional guidance compared with larger commercial coding assistants, particularly for complex programming tasks. The company recommends dividing larger projects into smaller instructions to improve results.
The startup previously secured a $410 million computing agreement with Amazon Web Services to support its research infrastructure. Socher said the long-term objective is to develop an automated research system capable of handling longer sequences of model-development tasks while allowing human researchers to supervise and intervene when necessary.
Featured image credits: Recursive Superintelligence
For more stories like it, click the +Follow button at the top of this page to follow us.