Recursive Releases Flint, a 40B Parameter Open-Weight Coding Model for Local Agents

Recursive Superintelligence has released Flint, its first open-weight coding model, according to MarkTechPost. The release targets security and engineering teams seeking local inference capabilities without cloud dependency. The model operates under an Apache 2.0 license, allowing organizations to inspect and modify weights stored on Hugging Face. This licensing choice affects how teams integrate the model into internal tooling and supply chain verification processes.

What are the technical specifications?

Flint is a decoder-only mixture-of-experts Transformer with 40 billion total parameters, but only 2.5 billion are active during processing, according to MarkTechPost. The architecture includes 49 layers, comprising two dense layers and 47 MoE layers. Each MoE layer contains 256 routed experts, with a router selecting six per token plus one shared expert. This design reduces computational load while maintaining performance on coding tasks.

View the original post on X

How does it perform on benchmarks?

The model scores 74.2 percent on SWE-bench Verified and 95.1 percent on IFEval, according to MarkTechPost. However, it trails dense models on Terminal-Bench 2.1, scoring 54.9 percent compared to 73.0 percent for Qwen 3.8. These self-reported scores used specific sampling parameters. Teams should validate these metrics in their own environments before trusting the model for critical code generation or vulnerability remediation tasks.

What hardware is required for deployment?

Flint runs on Apple Silicon Macs with 48 GB of memory or NVIDIA GPUs with similar capacity, according to MarkTechPost. The 4-bit quantized weights require approximately 23 GB of storage. Recursive developed specialized kernels for low-bit projections and MoE operations to optimize performance on M4 and M5 chips. This hardware requirement limits deployment to high-end consumer or workstation devices rather than standard servers.

How was the model developed?

Recursive states that its Eureka Machine, an automated research system, helped build nearly every part of the stack, according to MarkTechPost. The system proposed ideas, implemented them, and tested configurations. It started from a baseline of 35 billion total parameters and searched for optimal expert configurations. This automated approach influenced the final architecture and quantization strategy.

Questions to ask your suppliers

Organizations should ask vendors whether they audit models integrated into their products for known vulnerabilities. Teams usually verify that third-party AI components do not exfiltrate sensitive code or data. Check whether suppliers provide transparency into model training data and potential biases. Ensure that incident response plans cover scenarios involving compromised AI agents. Notify legal and compliance teams about any changes to data handling policies. Confirm that monitoring tools can detect anomalous behavior from autonomous coding agents.

Background: AI agents

AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. They move beyond generating text to performing tasks like coding, scheduling, or data analysis. You must treat them as active participants in your network, not passive tools.

Read the full guide: AI Agents Examples: How Autonomous Software Changes Security

What to do and how to stay safe: Recursive Superintelligence

• Monitor network traffic for unusual data exfiltration patterns from systems running the model.
• Verify that the model is isolated from sensitive internal repositories and production environments.
• Review logs for unexpected tool calls or file modifications made by the agent.
• Once the vendor provides an update, test it in a sandbox before wider deployment.

Step-by-step guide: Browser Security Settings for Small Business Teams

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What license does Flint use?

Flint is released under the Apache 2.0 license, according to MarkTechPost.

How many parameters are active in Flint?

Flint has 2.5 billion active parameters out of 40 billion total, according to MarkTechPost.

Can Flint run on standard laptops?

Flint requires 48 GB of memory, running on Apple Silicon Macs or NVIDIA GPUs, according to MarkTechPost.