Flint aims to accelerate local coding on computers

Recursive unveils Flint, a local coding model with a mixture-of-experts architecture. The tool targets agents working offline, with high inference speeds and open weights.

Recursive has introduced Flint, a local coding model designed to run agents on embedded AI hardware without relying on a remote API. This mixture-of-experts model has 40 billion parameters in total but activates only about 2.5 billion per token, reducing computational load during inference. The company says it built it in under four months with a small team of humans and agents, aided by its automated search system Eureka Machine. This topic matters to developers who want to keep data on their machine, limit costs, and work even offline.

A local model designed for busy developers

Flint mainly targets agentic coding, that is, assistants capable of planning and executing multiple steps of work on code. Recursive positions it as a fast model for local use, with announced performance of over 125 tokens per second on a MacBook Pro M4 Pro and 150 tokens per second on a 64 GB NVIDIA RTX Spark machine, with short contexts. The company also states that Flint supports up to 512,000 tokens of context, while recommending to stay under 256,000. A detail that will resonate with technical teams: the larger the context grows, the more models like to remind that they also have limits, with great confidence.

Very aggressive benchmark results, but to be read within their scope

Recursive claims that Flint leads MoE models under 50 billion parameters on SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified. The company also says its performance is comparable to dense models of 25 to 40 billion parameters, while running more than five times faster. These results come from the benchmarks cited by Recursive [to be verified by independent tests]. The company adds that Flint targets users who want to work locally, avoid API fees, and keep data on the device. In return, it recommends breaking down complex tasks into subproblems and giving explicit instructions, as the model may require more guidance than some broader closed alternatives.

Eureka Machine, the search engine behind the architecture

The development of Flint relies on Eureka Machine, Recursive’s automated system that proposed ideas, implemented components, and tested results. According to the company, this system influenced the architecture, data, optimizers, training recipes, infrastructure, quantization, and inference performance. The design path went from a 35 billion parameter base with 3 billion active to the final Flint configuration, which combines 40 billion parameters in total and 2.5 billion active per token. For quantization, Recursive says it placed the output layer at 4 bits and used offline distillation with the 16-bit model as teacher. The resulting NVFP4 model is within 0.7 points of the 16-bit version on internal benchmarks, while nearly halving training steps compared to its aware-quantization base.

Open weights, compatible tools, and dedicated inference engine

Recursive publishes Flint’s weights under the Apache 2.0 license, allowing broad reuse, including in commercial projects. The model is compatible with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang, facilitating adoption by teams already using local work tools. The company also launches rec, an inference engine tuned for Flint on Mac M4 Pro, M5 Pro, and M5 Max. In practice, Recursive sends a simple message: while the cloud retains the advantage for access to large generalist models, local computing is back in the race as soon as speed, privacy, and cost control are needed.

Key points

  • Flint activates 2.5 billion parameters out of 40 billion total.
  • Recursive built it in under four months.
  • 125+ tokens per second on MacBook Pro M4 Pro.
  • 150 tokens per second on NVIDIA RTX Spark 64 GB.
  • Apache 2.0 license for model weights.
  • Eureka Machine guided architecture, training, and inference.

By the numbers

  • 40 billion — total Flint parameters, according to Recursive.
  • 2.5 billion — active parameters per token, according to Recursive.
  • 512,000 tokens — maximum announced context window.
  • 0.7 point — maximum claimed gap with 16-bit version on internal benchmarks.
  • 4 months — announced development duration by Recursive.

To read