The Pulse
Recursive Releases Flint, a 40B Coding Model for Local Use
Recursive released Flint, a 40-billion-parameter coding model that activates 2.5 billion parameters per token and is designed to run on local hardware. The company says its Eureka Machine helped develop the model, whose weights are availabl
AI.info Team · Written with AI · fact-checked ·
Flint activates just 2.5 billion parameters out of its 40 billion for each token, a design Recursive says lets its new coding model run quickly on personal hardware. The company announced the model on October 21, presenting it as a local option for coding agents rather than a service that must send code to a remote server. Recursive says its small team built Flint from scratch in under four months, with help from its automated research system, the Eureka Machine.
Flint’s speed claims span laptops and server GPUs
Recursive reports speeds above 125 tokens per second on an M4 Pro MacBook Pro and above 150 on a 64GB NVIDIA RTX Spark, at short context lengths. The company says those rates reach as much as twice the speed of comparable models, and that Flint runs more than five times faster than dense models of 25 billion to 40 billion parameters with similar coding benchmark scores. Those are company-reported results from its optimized inference engine, not guarantees across hardware or software setups.
A separate launch-day integration post from vLLM, co-authored by Inferact and Recursive, reports about 500 tokens per second for a single request on an NVIDIA GB200 GPU, including with 100,000 tokens of context. That figure describes a server-class GPU, not a laptop, and reflects a different inference setup from Recursive’s Apple and RTX Spark numbers. The vLLM post says both Flint’s NVFP4 and BF16 checkpoints run on one GPU.
Benchmark lead comes with a disclosed test setup
Recursive says Flint leads mixture-of-experts models with fewer than 50 billion total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified. The company also says it rivals dense models in the 25-billion-to-40-billion-parameter range. Its evaluation notes describe agent runs with timeouts of up to five hours and averages across at least five attempts; SWE-Bench tests use mini-swe-agent without internet access, while Terminal-Bench uses the Terminus 2 harness with internet access.
The company’s own guidance is more qualified than a simple “runs locally” pitch: Flint may need more direction than larger closed models on complex work, and Recursive recommends breaking difficult tasks into smaller subtasks with explicit instructions. It lists a maximum context window of 512,000 tokens but recommends staying within 256,000 for better results. Those limits and the benchmark protocols matter to developers comparing results or fitting the model into an existing agent workflow.
The Eureka Machine helped build the model stack
Recursive says Eureka Machine contributed across Flint’s development, including data, training recipes, infrastructure and inference optimization. The company describes the system as a research loop that proposes ideas, implements them, tests results and uses those results to choose what to try next. Its account includes work on the model architecture, quantization and specialized kernels; the launch does not present Flint as a model produced without human involvement.
OpenHands, a launch partner, says it tested Flint ahead of release and supports it for software-engineering agents. “Recursive’s new model shows real promise for coding and agentic workloads, combining strong capabilities with the speed needed for practical software engineering tasks,” said Graham Neubig, co-founder and chief scientist at OpenHands. OpenHands also describes using Flint within an agent workflow that can inspect repositories, modify files, run tests and iterate on failures.
Weights are open; the evaluation question remains
Flint’s weights are available under the Apache 2.0 license, and Recursive lists support for llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. The company is also releasing rec, a lightweight inference engine optimized for Flint on M4 Pro and M5 Pro or Max chips. Developers can run the model on compatible local hardware, adapt it for their applications and keep code on their own devices.
Recursive’s published benchmarks and speed tests establish the company’s case for Flint, while the partner announcements show day-one integration with OpenHands and vLLM. Whether developers see similar results across their own hardware and long repository tasks is still an open test.