FAQ
Frequently Asked Questions
No. ISIRO is lossless compression. It is additive to quantization. It further compresses already-quantized models.
No. Weights are not approximated, retrained, or changed. Model output remains the same, at 100% model accuracy.
Decoded .tic weights match the source weights bit-for-bit. A hash manifest verifies that equality. KV cache compression (available for pilots) is also bit-for-bit exact.
Compiled .tic weights take less GPU memory, which leaves more room for KV cache. The KV cache is further compressed (available for pilots) for additional savings. Together that is more capacity on the hardware you already run.
BF16 is in general release. FP8 and FP4 are available for pilots.
See Models & Benchmarks for measured KV cache headroom and model footprint results, and the linked GitHub repo for memory, latency, and throughput.
Install for evaluation, and use the free compiler to build your own .tic. Then book a demo or pilot.