An open-weight model for locally programming agents, designed to work with Pi, OpenCode, and OpenHands for code understanding, modification, and testing.
Recursive FlintRecursive Flint is designed for developers who want to keep their code on their own devices: it handles the tasks of reading the code, generating modification proposals, and issuing tool calls, while programming frameworks such as Pi, OpenCode, and OpenHands handle file operations and testing. It is well-suited for using local AI for code interpretation, test completion, and small-scale fixes—rather than handing an entire software system over to a model and never checking it again.
Flint was developed by Recursive and is owned by Recursive Superintelligence, Inc. It uses a Mixed Expert (MoE) architecture with approximately 40 billion total parameters, activating about 2.5 billion parameters each time a token is generated. Reducing the number of parameters involved in a single step is a key aspect of its design for on-device inference, but this does not mean that only 2.5 billion parameters need to be loaded, nor does it mean that it can run on a typical low-memory laptop.
Recursive views the development of Flint as a practical application of Eureka Machine, its research system for automation. On October 21, 2026, OpenHands announced its support for Flint as a launch partner. The product is available as downloadable model weights and accompanying inference tools; the full development experience still depends on the programming framework used.Recursive Official Introduction - OpenHands Initial Support Announcement
Official Model Card:https://huggingface.co/RecursiveSI/flint
Flint itself outputs text and tool requests; the outer framework is responsible for executing commands. Connecting a terminal to the model does not automatically establish a security boundary; the working directory, network, keys, and executable operations must be controlled separately.Flint Capability Description
It is best suited for developers with prior experience with on-premises models, teams that need to control the flow of source code, and engineers looking to build self-hosted programming agents. If you simply want to install a ready-to-use chat app, or if your device has significantly limited memory, the effort required to set up Flint may outweigh the benefits.
| sports event | clarification |
|---|---|
| model size | 39.8B total parameters, approximately 2.5B activation parameters per token |
| Input and Output | Text; should not be used as an image recognition model |
| Inferential Behavior | Supports reasoning and tool invocation; for complex tasks, output budgets must be allocated for both the reasoning process and the final answer. |
| Context | The release notes state that it supports up to 512K and recommend keeping the size within 256K whenever possible; the public hosting configuration is set to 262,144 tokens, and the local limit depends on the runtime and available memory. |
| Weight Allowance | Apache 2.0; the terms governing the hosted API and data processing practices apply separately. |
Long context is a capacity metric; it does not guarantee that the system will accurately understand the entire repository if it is fed to it. Prompts, conversation history, tool responses, reasoning, and answers all consume context space.Model Card - Context and Runtime Settings
The most streamlined official method is using `rec` on a Mac, but its scope of support is fairly limited:
| Directions | Hardware and Platform Requirements | Please note |
|---|---|---|
| rec Local Quick Launch | M4 Pro, M5 Pro, or M5 Max Mac; at least 48 GB of unified memory; approximately 30 GB of available disk space; Node.js 22 and npm | For Macs with other chips or less than 48 GB of storage, rec will refuse to download the model; currently intended for single-session use |
| llama.cpp | Official support is available for Apple Silicon/Metal, as well as NVIDIA/CUDA on Linux or WSL2. | The Flint GGUF, runtime version, chat template, and tool parser must be matched; they cannot be mixed arbitrarily. |
| Original BF16 Weight Deployment | The SGLang example on the official model card uses a single B300 or H200, and the weights download to approximately 80 GB. | This is another server deployment option; it does not meet rec's space requirement of approximately 30 GB. |
Recursive has published local speed data under specific configurations, but these results cannot be directly extrapolated to other chips, quantization formats, or long conversations. It is particularly important to distinguish between “demonstrated RTX Spark performance” and “support for one-click installation via rec”: the current rec documentation still lists RTX Spark support as “coming soon.”rec Hardware Requirements - llama.cpp Path - BF16 Deployment Instructions
The following steps apply to Macs that meet the rec requirements:
The following are some commonly used commands:
rec doctor
rec launch pi
rec status
rec stopIf you're using a different client, you can start the service using `rec serve`. The official API endpoint is http://127.0.0.1:11500/v1; the model name is `flint`, and the local authentication token is provided by `rec auth token`. The token should only be placed in the client’s credentials field; do not include it in shared project configurations. Additionally, `localhost` in a container or remote development environment does not refer to this Mac.rec Installation and Connection Instructions
Flint weights are provided under the Apache 2.0 license, and the rec code is also licensed under Apache 2.0. There are no model API fees charged per token for downloading and running the model on your own, but there are still costs associated with hardware, electricity, storage, and maintenance. When using third-party hosting services, pricing, usage rates, and data terms are determined by the respective service providers; the availability of the weights should not be interpreted as meaning that all access is permanently free. The model and code licenses do not automatically cover the public reuse of images, logos, or other brand materials from the official website.Weighting and Licenses - rec Repository and License
Local inference can be performed offline once the model and dependencies are ready; however, the initial installation, download, and updates require access to the corresponding distribution sites. Download availability in different network environments and regional restrictions on third-party hosting services should be verified on a case-by-case basis; there is no single, universal regional commitment that applies to all Flint endpoints.
Labels and Licenses:The Flint product logo on this page is taken from RecursiveSI Official Model Repository...The original graphic is used for product identification only and is not used as a co-brand; it does not imply any collaboration or endorsement between the two parties. The warehouse includes... Full Text of the Apache 2.0 LicenseRetained for reference; use of the trademark remains subject to the limitations set forth in Article 6 of this license.
| offerings | Positioning and Functional Focus | More Suitable Tasks | Important Limitations |
|---|---|---|---|
| Recursive Flint | Approx. 40B / 2.5B local-programmed MoE per activation; features a dedicated rec route | For everyday code review, test coverage, and minor fixes on a Mac with plenty of RAM | Clear specifications for the REC chip and memory requirements; text input; single-session optimization |
| Qwen3-Coder-30B-A3B-Instruct | Qwen's Programming MoE, which supports calls from programming tools; native 256K context | I am currently using Qwen Code, Cline, or a general-purpose reasoning stack and would like to integrate an open-weight encoded model. | The function call format must be compatible; the extended context and quantization scheme will alter resource requirements, so the context limit should not be treated as a guarantee for low-end configurations. |
| Devstral Small 2 24B | Mistral's software engineering model supports multi-file editing and image input | Engineering tasks that require visual information, such as screenshots, in addition to code | It is essential to select a runtime stack that supports visual capabilities; long contexts and quantization configurations are still subject to device resource constraints. |
Both the official model cards for Qwen3-Coder-30B-A3B-Instruct and Devstral Small 2 are licensed under the Apache 2.0 License. This comparison focuses on deployment methods and task requirements; it does not rank products based on scores achieved under different frameworks or test configurations.Qwen Official Model Card - Mistral Official Model Card
Flint may generate error codes, misuse tools, or exhaust the output budget during lengthy reasoning processes. It is essential to maintain code reviews, testing, and permission isolation—in particular, do not allow initial trials to directly access production databases, production deployment permissions, or directories containing real keys. Complex requirements can be broken down into several phases—“describe the current state, propose a solution, conduct additional testing, make minimal changes, and verify the results”—which are easier to review than requiring a large-scale refactoring to be completed in a single go.Model Limitations
If your primary requirements are local control and reproducibility, and the tool meets those criteria, Flint is worth evaluating for your routine maintenance tasks. When deciding whether to use it long-term, prioritize your project’s repair success rate, test pass rates, and manual review time—rather than focusing solely on the number of tokens generated per second.
简体中文 简体中文 
English