RecursiveSI/flint is a compact, purpose-built model introduced alongside the Strata engine and Stratat client, which let you run capable AI locally instead of sending prompts to a cloud service. It is meant for on-device inference such as summarization, classification, or other short-generation tasks that require minimal latency and no internet connection after the app is installed.
Why RecursiveSI/flint exists
The goal of the model family is to bring useful generative intelligence to machines that cannot afford a datacenter GPU, by keeping all data on the device. Using Stratat, the model is loaded as a static page asset, and the Strata engine compiles the weights into an inference runtime that executes on the local CPU or accelerator, so user data never leaves the machine.
How it works
Strata provides a compilation path that converts the model’s weights into a format the device’s driver can execute. The Stratat client presents the model as a static page, allowing you to load the file, run a prompt, and see the generated text. Because the model is small, it fits within typical laptop memory and can be run on common x86 consumer CPUs.
| Metric | Intel Core i7-12700H test | AMD Ryzen 7 7840HS test |
|---|---|---|
| Peak RAM usage during inference | 4.8 GB | 5.1 GB |
| Average tokens per second | 11.2 t/s | 9.4 t/s |
| Power draw while running inference | 18.7 W | 22.3 W |
Power was measured with a plug-in wattmeter on the laptop's power brick; readings reflect the model running on the default Stratat power profile. Temperatures stayed under 75°C across the tests.
RecursiveSI/flint is a small, efficient state-interface model intended for local inference on laptops and edge devices through the Strata engine. Users can verify the model on their own machine by loading the static page in Stratat, entering a prompt, and observing the generated output. The following measurements were taken on two reference laptops using the same model version: the Intel Core i7-12700H system reported approximately 4.8 GB of RAM usage during inference, completed about 11.2 tokens per second, and drew about 18.7 W while running, while the AMD Ryzen 7 7840HS system reported about 5.1 GB of RAM usage during inference, completed about 9.4 tokens per second, and drew about 22.3 W while running. To check these numbers yourself, open the RecursiveSI/flint static page in Stratat, select the model from the drop-down, enter a short prompt, and note the displayed token rate and load indicators. Because the model is small enough to fit in typical laptop memory, it can run on many consumer CPUs without a dedicated accelerator.
What the core term means
The core term is state-interface model: a lightweight neural model that maintains a compact internal representation (a state) of recent context and generates text token by token, optimized for low latency and low memory usage on local devices rather than for maximum throughput on server-class hardware.
Measuring performance on your own machine
To reproduce the measurements, open the Stratat client and navigate to the RecursiveSI/flint static page. Load the model, choose a short prompt, and run inference. Record RAM usage shown in the status bar, the tokens generated per second, and the device’s power draw using a standard plug-in wattmeter. Compare these values with the figures above to confirm whether your hardware matches the reported range.
Key takeaways
RecursiveSI/flint is intended for private, on-device AI tasks, and it runs entirely on the machine without contacting external services. Its small size and optimized inference make it practical for everyday laptops using the Stratat desktop app.