Recursive releases Flint, a 40B open-weight coding model, with day-0 support from inference tools
Article timeline
The articles in this story. Times are ET.
ollamaOfficialPickOllama announces Flint, a ~40B MoE open coding model from Recursive_SIAIOllama introduces Flint, a ~40B-parameter MoE open model from Recursive_SI for agentic coding on local hardware. The post says it generates over 125 tokens per second on an M4 Pro MacBook Pro and over 150 on a 64GB NVIDIA RTX Spark at short contexts, and users can run it with ollama run flint or ollama launch claude --model flint.
Sophia YangXPickRecursive's Flint is an open-weights local coding model built in under four monthsAIRecursive, through Sophia Yang's post, introduces Flint, its first model, a fast local coding model built with its Eureka Machine automated research system. The model is a 40B total, 2.5B active MoE with Apache 2.0 weights on Hugging Face, and it scores 54.9 on Terminal-Bench 2.1 versus 44.9 for Qwen3.6 35B-A3B. Dense Qwen3.8 27B still leads on coding with a Terminal-Bench score of 73.0, and Flint runs at 125+ tokens per second on an M4 Pro MacBook at short context.
Lewis TunstallXLewis Tunstall praises Recursive's Flint, a sub-100B coding modelAIHugging Face's Lewis Tunstall praises Recursive's release of Flint, calling it proof that sub-100B models can still innovate. He asks to see the time-horizon chart for Flint's training by Eureka, the Eureka Machine, citing an OpenAI image.
SGLangOfficialSGLang adds Day-0 support for Recursive's Flint local coding modelAISGLang says it has Day-0 support for Flint, a 40B MoE coding model from Recursive, which the company says leads its class on SWE-Bench Pro v2 and Terminal-Bench 2.1. The post gives a sglang serve command for running Flint locally on DGX Spark with a 262144-token context length, and the model's open weights are available on Hugging Face.
LM StudioOfficialPickLM Studio adds Flint, a local coding model from Recursive, on day 0AILM Studio says Flint, the first model from Recursive, is now available in LM Studio on release day. Flint is a 40B mixture-of-experts model with 2.5B active parameters, a 262K token context window, and an Apache 2.0 license. The post says it needs about 24GB to run locally and claims up to 2x faster speeds than comparable models, at 125+ tokens/sec on a MacBook Pro and 150+ on RTX Spark.
vLLMOfficialPickvLLM adds day-0 support for Recursive's 40B Flint coding modelAIvLLM says it has day-0 support for Recursive's Flint, a 40B-parameter model with 2.5B active parameters per token. With custom NVFP4 kernels, it reaches 500 tokens per second on a single GB200 at 100K context without speculative decoding. NVFP4 and BF16 checkpoints are both available now.
elvisXRecursive SI releases Flint, a 40B open-weight coding model built with automated researchAIRecursive SI releases Flint, a 40B mixture-of-experts model with 2.5B active parameters and open weights under Apache 2.0. Its Eureka Machine contributed to nearly every part of the stack, from data and training recipes to inference kernels, and the team built the model in under four months. Flint runs at 125+ tokens per second on an M4 Pro MacBook Pro at short contexts.
SantiagoXFlint, a 40B open-weight local coding model, runs at 125 tokens/sec on M4 MacBook ProAISantiago highlights Flint, an open-weight coding model that runs locally on regular computers. Flint is a 40B-parameter mixture-of-experts model that activates 2.5B parameters per generated token and has a 512K context window. He reports 125 tokens/sec on an M4 MacBook Pro.
- Wired · AINewsPickRecursive launches Flint, an open-weight coding model its CEO says AI helped build
AIRecursive has released Flint, an open-weight AI coding model with 40 billion parameters, which CEO Richard Socher says was built largely by its Eureka Machine using recursive self-improvement. Wired's Will Knight tested Flint on a Mac with 48 GB of RAM and found it fast and capable on his coding requests. Recursive has raised $650 million at a $4.65 billion valuation, and Socher says the Eureka Machine is designed to remain under human oversight.
- vLLM BlogOfficialvLLM adds day-0 support for Recursive's Flint model on a single GPU
AIvLLM supports Recursive's Flint from day 0, and both its NVFP4 and BF16 checkpoints run on a single GB200 GPU. A single request decodes at about 500 tokens/s with 100K tokens of context. The post lists serving commands and details the kernel changes that cut NVFP4 decode time from 4.81 ms to 1.78 ms per step.