Strata supports DeepSeek-AI DeepInfer for local inference on large models

DeepSeek-AI released the DeepInfer model for structured reasoning tasks, and Stratat now includes a ready-made client that can execute it locally on consumer GPUs. This page explains what DeepInfer does, what hardware the Stratat client needs, and how you can verify the performance on your own machine.

What DeepInfer is and why it matters

deepseek-ai/DeepInfer introduced a model family designed for efficient, deterministic reasoning over long sequences of text. Stratat adds a client that packages the model and its runtime so it can be launched from a desktop page without any network call to DeepSeek's servers. Using DeepInfer locally means your prompts stay on your machine, which helps keep sensitive context private.

Hardware requirements for running DeepInfer with Stratat

the Stratat client expects a system that has a recent NVIDIA GPU with at least 8 GB of VRAM and at least 16 GB of system RAM. The engine reads model weights from disk, so a fast NVMe drive is recommended. If you have less than 12 GB of VRAM the model will not load, and an Intel CPU without a CUDA-capable GPU cannot run the deepseek-ai/DeepInfer path.

How to test DeepInfer yourself

open a Stratat window and load the DeepInfer checkpoint. Run the provided timing script on a short prompt and note the seconds for one forward pass. Compare those timings against the baseline numbers shown in the table below. You can also monitor GPU memory usage via the built-in diagnostics pane while the model is generating.

DeepInfer is defined as "deterministic chain-of-thought prompting that runs fully on-device to preserve privacy".

Checking the results on your own hardware

run the script `strat-test.sh` inside the DeepInfer example folder. it prints the elapsed time per generation step and the peak GPU memory. compare those values with the table above to confirm that the model behaves as advertised on your specific machine.