Recursive released Flint on October 21, 2026, as a coding model intended for local agent workflows, according to TechCrunch’s release report. Recursive’s model card describes 39.8 billion total parameters and approximately 2.5 billion active for each token. That active figure describes part of the computation. It does not mean the complete model occupies the space of a 2.5-billion-parameter model.
For a developer deciding whether to run Flint locally, the consequential choice is the checkpoint and runtime.
The checkpoint matters more than the active count
The model card’s main BF16 checkpoint is about 80GB to download. Its documented SGLang serving route uses a B300 or H200 GPU. The card says quantized variants for Apple Silicon and NVIDIA runtimes are prepared separately. Downloading the BF16 repository is therefore different from choosing a laptop-ready build.
Recursive’s rec inference-engine documentation describes a Mac route using 4-bit weights. It specifies an M4 Pro, M5 Pro, or M5 Max Mac with at least 48GB of unified memory and about 30GB of free disk space. The engine is optimized for a single session and can connect to Pi, OpenCode, or a local API.
Those requirements apply to the documented rec setup. They are not a universal minimum for every Flint quantization or runtime. Developers considering another route should check that build’s support and resource requirements rather than infer them from the 2.5B active-parameter figure.
Context and benchmarks need qualification
Recursive’s launch announcement advertises context windows of up to 512,000 tokens, while recommending workloads within 256,000 for best results. The model card documents a 262,144-token configuration that includes both input and output. The pages do not establish that every available runtime supports the larger advertised maximum, and the card warns that capacity does not guarantee consistent accuracy across long sequences.
Recursive also reports coding-benchmark and throughput results from its own evaluations. The model card notes that some comparison scores are developer-reported and evaluation setups differ. Those numbers are useful starting points for testing, not an independently established ranking across coding assistants. Recursive itself says complex tasks may require more explicit instructions than larger commercial models.
What to test before relying on it
A practical trial should use the exact checkpoint and runtime intended for deployment, with representative repository tasks and context lengths. Confirm the machine’s memory and storage first, then assess code quality, latency, and tool behavior in that setup.
The model card’s limitations warn that Flint can produce faulty code or inappropriate tool calls. Generated changes still need review and tests, while tool-using applications need their own permission and execution boundaries. Local inference can avoid sending prompts to a hosted model API, but teams should inspect their chosen harness and integrations before assuming the entire workflow stays on the device.
