Strata 0.2.9 has been released. I ran a quick benchmark, but it doesn’t seem to differ from the previous version. Using an RTX 5070 Ti and 64GB of RAM with the Qwen 3.8FN IQ3_S model, I’m seeing an average decode speed of 85 t/s, with prefill speeds occasionally exceeding 2,700 t/s. Even with a context length of over 120K—pushing toward 150K tokens—the loading is so fast that the difference is clearly noticeable in daily use. That said, given my memory constraints, running any larger models isn’t an option.
