Flint 40B-A2.5B

Flint released a 40B-parameter mixture-of-experts coding model, Flint 40B-A2.5B, with weights on HuggingFace under Apache-2.0 including official INT4 and NVFP4 GGUFs, designed to run agentic coding entirely on local machines such as a 48 GB laptop or an RTX Spark. The model activates only 2.5B parameters per token, supports a 128k context, and runs at 131 words per moment on an M4 Pro, with vendor-measured benchmarks of 74.2 on SWE-bench Verified in BF16 and 54.9 on Terminal-Bench 2.1. Flint positions the release as breaking the trade-off between API dependence and models too slow for useful agentic coding, with results reproducible locally.

Flint 40B-A2.5B MoE workstation Flint is a coding model with one job: doing real work in real repos, entirely on your own machine. Think of it as a senior contractor who works at 131 words per moment on an M4 Pro and never phones home. It exists because agentic coding was stuck between API dependence and models too slow to be useful; the MoE design 40B total, only 2.5B read per token breaks that trade. Reach for it when the work is bug hunts, tests, feature code, and the machine is a 48 GB laptop or an RTX Spark. Honest status: weights on HuggingFace Apache-2.0, official INT4 and NVFP4 GGUFs , benchmarks vendor-measured SWE-bench Verified 74.2 BF16, Terminal-Bench 2.1 54.9 , reproducible at home tonight. coding agentic local moe apple-silicon nvidia - Parameters - 40.0B - Context - 128k - License - apache 2.0 - Origin - 🇺🇸 USA - Released - Oct 2026 PRICE HISTORY Inference cost over time Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API. Loading price history...