Recursive Flint

Recursive Flint is Recursive's Apache-2.0 open-weight mixture-of-experts model for coding agents. The official model card describes 39.8 billion total parameters and about 2.5 billion active parameters per token. It is designed for codebase investigation, bug fixes, feature work, test writing, multi-turn reasoning, and tool calls rather than as a general hosted-chat endpoint.

What the release includes

• A local-serving package: The published repository contains a BF16 checkpoint, tokenizer, chat template, model configuration, and an SGLang serving kit. Recursive documents OpenAI-compatible local requests after the server is running.
• Sparse architecture, real hardware needs: The rounded “40B-A2.5B” label describes 39.8B total and roughly 2.5B active parameters. Active parameters reduce work per token; they do not make the full checkpoint or its memory requirements disappear.
• Agent-focused interface: Flint supports reasoning and OpenAI-compatible tool calls. Preserve reasoning_content in the conversation history and use the supplied chat template, as instructed by the model card.
• Documented capacity: Its configuration lists 262,144 tokens shared by prompt and generated output. Treat that as a capacity boundary, not evidence that every long context will be equally accurate.

A practical evaluation workflow

Start with a reproducible issue or a small repository and grant read-only tools first. Downloading the checkpoint is documented as roughly 80 GB, and Recursive documents one B300 or H200 GPU per SGLang replica. Run the provided health check, then compare Flint and a smaller model on the same task with fixed tool permissions, time limit, output budget, and test suite. Log the tool trace as well as whether the proposed patch passes tests; an agent benchmark score alone cannot establish reliability on your codebase.

For a first integration, put the SGLang endpoint behind a local network boundary, retain only the request data your policy permits, and require review before writes, commands, or deployments. If a managed provider offers Flint later, verify its separate pricing, data-retention, region, and terms instead of assuming that the Apache-2.0 weight license covers the service.

Limits and alternatives

Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified under its stated setups. These are vendor-reported evaluation results, and the card notes that setup differences can affect comparisons. Long reasoning can add latency and consume the output budget; tool-call structure does not enforce safe permissions. Test generated code before applying it.

For a smaller open coding-agent model, compare LFM2.5-24B-A2B. For a much larger open-weight option built for long-horizon agent work, compare Reflection Beam 1.0. Read the official model card and Recursive launch post before provisioning hardware.