Recursive's Flint: An AI-Built Coding Model Designed for Local Hardware and Top Benchmarks

AI startup Recursive has released Flint, its first language model, a 40-billion-parameter mixture-of-experts (MoE) model largely developed by its own AI research system, the "Eureka Machine." Designed for agentic coding on local hardware, Flint claims to lead coding benchmarks among MoE models under 50 billion parameters and can generate over 125 tokens per second on an M4 Pro MacBook. For broader context, explore our Top 100 AI Tools.

Flint's AI-Driven Development and Architecture

Recursive's Flint model represents a notable step in AI-assisted AI development. The company states that its "Eureka Machine" AI system played a role in nearly every stage of Flint's creation, a process completed within four months. This approach highlights a growing trend where AI tools contribute to their own evolution.

Architecturally, Flint is a 40-billion-parameter mixture-of-experts model, utilizing 2.5 billion active parameters per token. This design choice is intended to balance performance with efficiency, making it suitable for deployment on local hardware. The model's weights are openly available under the Apache 2.0 license, promoting accessibility and further development within the AI community.

Performance Claims and Local Hardware Compatibility

Recursive has positioned Flint as a high-performance solution for agentic coding. The company reports that Flint achieves over 125 tokens per second on an M4 Pro MacBook equipped with 48 GB of memory. This performance metric suggests that Flint is optimized for efficient operation on Apple Silicon, potentially reducing the need for extensive cloud-based computational resources for certain coding tasks.

Flint's compatibility extends to several popular inference engines, including llama.cpp, Ollama, LM Studio, vLLM, and SGLang. This broad support aims to facilitate integration into existing developer workflows and environments. The model also features a substantial 512k-token context window, with a recommended 256k for practical use, allowing it to handle complex and extensive codebases.

Benchmark Performance and Verification Status

Recursive asserts that Flint leads in key coding benchmarks such as SWE-Bench Verified and Terminal-Bench 2.1. These claims specifically apply to MoE models with fewer than 50 billion parameters. While these results indicate strong performance, Recursive notes that the benchmark scores have not yet undergone independent verification. This distinction is important for evaluating the model's capabilities within the broader landscape of AI news and research.

Implications for Local AI Development

The release of Flint underscores a strategic focus on local-first AI models, particularly for specialized tasks like coding. By enabling high-speed inference on consumer hardware, Recursive aims to empower developers with powerful AI tools that can operate without constant internet connectivity or reliance on remote servers. This local execution capability can enhance privacy, reduce latency, and potentially lower operational costs for individual developers and small teams.

The open-source nature of Flint's weights under the Apache 2.0 license further supports its adoption and integration into various projects. This licensing model encourages community contributions and allows for customization, which could accelerate the development of new AI-powered coding agents and tools.

Conclusion

Recursive's Flint model, developed with significant AI assistance from its "Eureka Machine," represents a notable entry into the field of local-first coding AI. With its claimed benchmark leadership among MoE models under 50 billion parameters and efficient performance on devices like the M4 Pro MacBook, Flint offers a compelling option for developers seeking powerful, locally executable coding assistance. As the AI community continues to evaluate and independently verify its performance, Flint's open-source availability and local hardware optimization position it as a model to watch in the evolving landscape of AI-driven development tools.

Sources

• RecursiveSI/flint · Hugging Face
• GitHub - recursive-oss/rec: Our lightning-fast inference engine. · GitHub
• OsaurusAI/Flint-40B-A2.5B-JANGH2 · Hugging Face
• EigenLabs/flint-int4 · Hugging Face
• Recursive launches Flint, a laptop coding model it says was largely built by AI | TechCrunch