Flint is Recursive’s text-only coding model, with 39.8B total parameters and about 2.5B active per token. This estimate uses Ollama’s 23 GB compressed download, rather than the roughly 80 GB publisher BF16 checkpoint.
Where it works well
- Coding-focused reasoning.
- Multi-turn tool use in a connected app.
- Local quantized inference.
Common use cases
- Explain a reproducible bug.
- Propose a focused refactor.
- Draft tests for an edge case.
Know its limits
- All weights need memory; 2.5B active does not make it a small model.
- It cannot see screenshots; provide text errors and code.
- File editing and command execution require a coding app. Review and test generated changes.
Run it on your computer
Start with a compressed download and a short conversation. Our estimate is about 48 GB of total RAM or Mac unified memory, with room for the operating system. The listed 23 GB download is the model file, not its full running memory use.
Check the model’s supported runtime and hardware before loading it. System RAM on a PC and unified memory on a Mac are different from dedicated NVIDIA VRAM. A memory match doesn’t guarantee compatibility or speed. Choose a smaller model if responses are too slow.
Use a desktop app that supports the listed model format, such as LM Studio or Ollama. Download the exact local variant, load it in Chat, and start with one small task. Model downloads need internet; local chat can work offline once the model is downloaded.
Official model information
Model capabilities checked October 2026. The use cases above are practical starting suggestions, not performance guarantees.