Recursive releases rec for local Flint inference on Macs
Recursive’s rec v0.1.0 release is dated October 21 on GitHub. That record establishes the engine’s release day. It does not establish when Flint’s model weights first became public.
The company’s Flint announcement describes a coding model with roughly 40 billion parameters, activating about 2.5 billion per token. Recursive says its automated research system helped develop the architecture, training and inference stack. Those are the developer’s claims.
Which Macs can run rec
The engine documentation requires an M4 Pro, M5 Pro or M5 Max Mac, at least 48 GB of unified memory and about 30 GB of disk space. The pi recipe additionally needs Node.js 22 and npm.
rec is optimized for one session. It provides pi and OpenCode recipes and an API compatible with OpenAI chat completions. First launch downloads the model and harness. The model server keeps running after the terminal closes, until stopped with rec stop.
Recursive reports 1.4 to 2.6 times llama.cpp’s token throughput in its specified M4 Pro test. That comparison uses 48 GB memory, four bit weights, five runs per context and excludes prompt processing. It is a vendor decode benchmark.
Two different deployment paths
The main Flint model card distributes BF16 weights, an approximately 80 GB download, with an SGLang kit for one B300 or H200 GPU per replica. It says Transformers modeling code is not yet included.
The model card identifies Apache 2.0 for the weights and separates hosted API terms from that license.
The rec software repository has its own Apache 2.0 license. That finding concerns the engine code.
The context figures need qualification
Recursive’s announcement claims support up to 512k tokens and recommends staying within 256k. The model card instead documents a 262,144 token configuration shared by prompt and output. These are different published specifications, not a verified 512k setting for every runtime.
For complex coding work, the announcement recommends smaller subtasks and explicit instructions. Recursive acknowledges that Flint can require more guidance than larger closed models. Its automated research narrative does not replace testing on a developer’s own workload.
Model scores need context
The model card reports agentic scores averaged across at least five attempts and warns that evaluation setups differ. Its Terminal Bench result treats solutions obtained through web retrieval as unsuccessful, while comparison scores may use different rules. These are reported evaluations.
The card also warns about faulty code and inappropriate tool calls. Applications must enforce their own permissions. Structured tool calling does not prevent an unsafe choice.
ByteForward reviewed the published materials without installing rec or running model benchmarks.
Related coverage examines Volundr and Agens Engine deployment limits.
Illustrative archival photograph by Mahmudul Hasan via Unsplash under the Unsplash License.
