Recursive Flint Puts a Fast Agentic Coding Model on Developers’ Laptops

Recursive Superintelligence, the AI lab led by former Salesforce chief scientist Richard Socher, has released its first model. Flint is an open-weight coding model built to run locally rather than through a cloud API.

That matters for DevOps teams. Many organizations want AI coding help but can’t, or won’t, send source code to a third-party service. Flint is aimed at that gap.

What Flint Is

Flint is a mixture-of-experts (MoE) model with 40 billion total parameters. It uses only about 2.5 billion of them for each token. That keeps per-token compute low so that it can run fast on a laptop or desktop.

Recursive says Flint leads MoE models under 50 billion total parameters on agentic coding benchmarks, including SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. The company also says it matches dense models in the 25- to 40-billion-parameter range while running more than five times faster.

On speed, Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro. It reports more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short context lengths.

These are Recursive’s own numbers. No one has verified them independently yet. Teams should test Flint on their own codebases before concluding.

Flint supports context windows up to 512,000 tokens, but Recursive recommends staying under 256,000 for best results. The company recommends 48GB of memory for agentic coding work. LM Studio lists a quantized build that needs at least 24GB.

How Teams Can Run It

The weights are on Hugging Face under the Apache 2.0 license. Teams can download, modify, and deploy the model commercially without licensing fees.

Flint works with common local inference tools, including llama.cpp, Ollama, LM Studio, Unsloth, vLLM, and SGLang. Recursive also released rec, a lightweight inference engine tuned for Flint on Apple’s M4 Pro and M5 Pro/Max chips. OpenHands, the open-source coding agent platform, signed on as a day-zero launch partner. A free hosted version is also on OpenRouter for teams that want to try it before setting up hardware.

Recursive is upfront about the limits. The company says Flint may need more guidance than the large closed-source coding models developers are used to. It advises breaking complex work into smaller subtasks and giving explicit instructions. That’s good practice with any coding agent. But it matters more with a small model.

Built by an AI Research System

The more unusual part of the story is how Flint was made. Recursive says a small team of people and AI agents built it from scratch in under four months. The company credits its Eureka Machine, an automated research system that proposes ideas, implements them, tests them, and uses the results to decide what to try next.

Recursive says the system contributed to nearly every layer of the model-building stack. Three examples stand out.

Architecture. Starting from a baseline of 35 billion total and 3 billion active parameters, the system tested expert configurations. It landed on 40 billion total with 2.5 billion active as the best balance of speed and coding performance.

Quantization. The system compressed the model’s output layer to 4 bits, which goes against common practice. To recover accuracy, it added a step where the compressed model trains on its own pre-generated responses, with the full-precision model as teacher. Recursive says the final NVFP4 model scores within 0.7 points of the 16-bit version.

Kernels. The system wrote specialized kernels for Apple Silicon and NVIDIA RTX Spark. On M4 and M5 chips, that includes a single Metal kernel that handles the entire decode pass.

Recursive came out of stealth in May with $650 million in funding at a $4.65 billion valuation, in a round co-led by GV. Its stated goal is self-improving AI that automates knowledge discovery. Flint is the first public test of that idea.

Why It Matters for DevOps

Local coding models change a few things for platform and DevOps teams.

The first is cost. There are no per-token API fees. The cost is hardware, which many developers already have.

The second is control, and it cuts both ways.

“Local models remove a major barrier to AI coding agents, source code leaving the company,” said Mitch Ashley, vice president and practice lead for CIO & Technology Buyers and Software Lifecycle Engineering at The Futurum Group. “They also put agent governance on every developer’s laptop, outside the platform team’s view. Teams adopting them need an approved model list and a record of what each agent changed.”

And the review question doesn’t go away. A faster model writes more code, faster. That code still has to pass testing, security scanning, and human review before it reaches production. Whoever deploys an agent owns what it produces.

What Comes Next

Recursive says more capable models should make the Eureka Machine a better researcher, speeding up each new generation. The company plans more local models and the infrastructure to run them.

Ashley sees that pace as a challenge in its own right.

“If Recursive’s automated research system can build a competitive model in under four months, new models will arrive faster than teams can evaluate them,” he said. “Evaluation becomes continuous work, tested against each team’s own code before a model reaches the pipeline.”

Whether Recursive’s loop holds up is an open question. For now, Flint gives DevOps teams another open-weight option for local agentic coding. It’s worth a test run, especially for teams that have held off on AI coding tools because of where their code would end up.