Recursive released Flint on October 21, giving developers downloadable weights for a coding model and a dedicated Mac inference engine called rec. The launch announcement pitches local coding without sending data to a remote model server. Its recommended Mac route has a practical boundary: rec handles one generation at a time, and overlapping requests can receive HTTP 429, according to the setup documentation.
That makes the release most immediately relevant to a developer working with one interactive coding agent. If you want several agents generating answers concurrently, you'll need to assess a different serving setup. The rec repository describes the engine as optimized for single-session use and directs users elsewhere for multiple agents. MrKeyoor has not run Flint or reproduced Recursive's performance results; this report examines the release and its documented deployment requirements.
The Mac route has a hardware check
The rec instructions list an M4 Pro, M5 Pro or M5 Max Mac with Metal support, at least 48GB of unified memory and about 30GB of free disk space. The launcher checks the chip and memory before downloading the model. Node.js 22 and npm are also required to launch the Pi or OpenCode coding harness, the application that handles the model's interaction with your project.
Once installed, rec can start its local server and connect either harness automatically. The first launch downloads the model and the matching harness version. Closing the terminal doesn't stop the model server. The repository's usage instructions say to run rec stop when you're finished.
The standard checkpoint is a different deployment
Flint's model card specifies 39.8 billion total parameters, with approximately 2.5 billion active per token. It uses a mixture-of-experts architecture, activating part of the network for each token. The downloadable repository contains BF16 weights; quantized versions are prepared separately for Apple Silicon and NVIDIA runtimes.
The card's SGLang recipe calls for about 80GB of files and one B300 or H200 GPU per replica. Those requirements describe the BF16 serving path, rather than the Mac package's roughly 30GB disk allowance. Flint is released under the Apache 2.0 license. The card says hosted API terms and data practices are separate from the weight license.
A coding score needs its serving context
Recursive reports 74.2% on SWE-bench Verified in its model card. Its agentic results use average pass@1 over at least five attempts, with SWE-bench run through mini-swe-agent. The card warns that evaluation setups vary and that results can change with checkpoint precision and inference runtime. These are Recursive's measurements, so they don't establish the success rate of the compressed Mac version on your repository.
The company also attributes much of Flint's development to its Eureka Machine automated research system. Its announcement describes contributions to architecture selection, quantization and execution kernels, while advising users to divide complex coding work into smaller tasks and provide explicit instructions. It acknowledges that Flint may need more guidance than larger closed models.
The next useful evidence is completed coding work on the released local configurations: which changes pass tests, how much steering they need and how long they take. Recursive's rec documentation says support for NVIDIA RTX Spark is coming soon. Its automatic setup instructions currently apply to supported Macs.