# Recursive Brings Flint Coding Model to Local Machines With 2.5B Active Parameters By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/37525 Published: 2026-10-22T05:35:40.000Z Updated: 2026-10-22T05:35:40.000Z Section: Technology > The open-weight model has 39.8 billion parameters in total. Its BF16 checkpoint is about 80GB, and its Apple Silicon setup requires at least 48GB of unified memory. Recursive released Flint on Oct. 21 as an open-weight model for coding agents that can run on local machines. Its mixture-of-experts design activates about 2.5 billion of its 39.8 billion parameters for each token, and its weights are available under the Apache 2.0 license. A mixture-of-experts model uses only part of its parameter set for each token. Flint’s active parameter count is a subset of its total, which lets the model use fewer parameters per token than its full size might suggest. At short context lengths, Flint produces more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark. Those figures apply to short contexts, where the model has less text to process. Running Flint through Recursive’s Rec engine on Apple Silicon requires a Mac with an M4 Pro, M5 Pro or M5 Max chip, at least 48GB of unified memory and about 30GB of free disk space. The BF16 checkpoint is about 80GB to download. Separate quantized versions are prepared for NVIDIA and Apple Silicon runtimes. Flint scored 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. The agentic scores average pass@1 across at least five attempts. Evaluation setups differ, so the scores reflect results under varying conditions. A small team of people and AI agents built Flint from scratch in under four months using Recursive’s Eureka Machine. “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” said Richard Socher, Recursive’s CEO and co-founder. Running Flint locally can keep code on a developer’s machine without sending each request to a remote API. Flint’s memory and storage requirements mean local use calls for capable hardware. TokenPost previously covered [NVIDIA’s addition of Flint to the RTX Spark lineup](). ## Links in this article - [NVIDIA’s addition of Flint to the RTX Spark lineup](https://www.tokenpost.com/news/technology/37449)