Skip to main content
Flint is designed for agentic coding: understanding repositories, generating code, making edits, and running tests. The model generates text and tool requests; your harness executes the tools and returns their results. See the model card for architecture, checkpoint formats, and license.

​
Run on your Mac

Use rec for Recursive’s optimized Mac inference engine and automatic Pi and OpenCode setup. You need an M4 Pro, M5 Pro, or M5 Max with at least 48 GB of unified memory. After installing rec, launch from your project directory:
rec starts the model server and configures Pi, our recommended local harness. For OpenCode, use rec launch opencode instead. To use your own serving stack, see local serving workflows.

​
Capabilities

​
Context length

The hosted Flint configuration uses a 262,144-token context window, with an output budget of up to 65,536 tokens. Context includes the prompt, conversation history, tool definitions, tool results, reasoning, and the answer. For local inference, use the limit configured by your runtime. Available memory, quantization, and cache settings determine how much context fits on your Mac; the hosted limit is not a local memory guarantee. With rec, keep the context and output settings supplied by its launcher. With llama.cpp, set --ctx-size and match your harness to that value.

​
Reasoning and generation settings

Allow enough output tokens for both reasoning and the final answer. If a response ends with finish_reason: length, increase the output budget within your runtime’s limit or shorten the conversation. Use your runtime’s generation settings. When launching Pi or OpenCode with rec, keep the launcher’s bundled sampling and output settings.