Recursive released Flint on October 21, giving developers a new coding model they can download and run on their own hardware. It targets repository work—investigating bugs, implementing features, and writing tests—with reasoning and multi-turn tool use rather than autocomplete alone.
Open weights and repository-sized context
- Open weights: Apache 2.0 licensing, with 39.8 billion total parameters and approximately 2.5 billion active per token.
- Context configuration: the published model card documents 262,144 tokens shared between input and output.
- Deployment choice: support includes Ollama, llama.cpp, LM Studio, vLLM, and SGLang.
The active-parameter count describes per-token computation, not the full model’s storage footprint. Teams should evaluate the checkpoint and runtime they intend to use.
Local speed comes with hardware requirements
Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts using its optimized engine. That is a vendor measurement, not a guarantee for every workload or context length.
Its rec engine currently requires an M4 Pro, M5 Pro, or M5 Max Mac with at least 48 GB of unified memory and is optimized for a single session. Those are requirements for that engine, not a claim that every supported Flint deployment uses the same hardware.
Why it matters for builders
Local coding agents offer an alternative for teams wanting to keep model inference and source code on-device and avoid per-token hosted API charges. Hardware, setup time, and review effort still contribute to the total cost.
For SunMarc App Labs, Flint is worth evaluating on bounded bug fixes and test-writing tasks, measuring completed work and review effort—not just generation speed. This is a proposed evaluation, not a claim that SunMarc has tested or deployed Flint.
A concrete automated-research result
The release makes Recursive’s automated-research pitch more concrete: the company says its Eureka Machine helped develop Flint’s architecture, training recipes, and inference stack. Developers now have a downloadable model to evaluate, rather than only a research promise.