Recursive SI Releases Flint Open-Weights Model for Fast Local Coding
Recursive SI unveiled Flint, its first open-weights coding model built in four months using Eureka Machine, targeting high-speed local inference on consumer har
Image source: Richard Socher (@RichardSocher) / Recursive SI
Automated Creation via Eureka Machine
Recursive SI has introduced Flint, its debut model engineered specifically for fast local coding. Developed from scratch in under four months, Flint was built with the company's automated research system called the Eureka Machine. The announcement describes the model as providing private, unmetered intelligence on a user's own machine. Richard Socher attributes the development timeline of less than four months in part to the Eureka Machine.
Execution Speed and Model Distribution
According to the release, Flint targets rapid throughput on personal hardware. Reported generation speeds reach 125+ tokens per second on a MacBook Pro and 150+ tokens per second on RTX Spark hardware, which Recursive SI claims is up to 2x faster than comparable models. Open weights have been made available on Hugging Face for local runtime deployment across supported frameworks. The announcement also points to free trials via @opencode or @pidotdev.
Evaluation Gaps and Vendor Benchmarks
While the reported local generation speeds are notable, the 125+ and 150+ tokens per second throughput figures remain self-reported vendor numbers rather than independently audited metrics. The supplied evidence contains no independent verification across diverse environments or configuration details, such as specific quantization parameters and operational setups, to confirm real-world performance ceilings. Additionally, the primary announcement did not disclose the model's exact parameter size or formal benchmark evaluation scores, leaving code accuracy and comprehensive task completion to be evaluated as further technical documentation becomes available.