All news

SiMa.ai speeds up SI models on Modalix chips by up to 3.5x

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Device developers on Modalix chips get up to 3.5x faster token generation without higher power consumption, and the open code lets them verify the claimed figures.

SiMa.ai has released Palette 3.0, a software update for Modalix edge chips. The update was announced on November 17. It speeds up token generation by up to 3.5x within the same power budget thanks to speculative decoding with the EAGLE3, DFlash and Gemma 4 MTP algorithms. In the company's white paper, the Qwen3.5 9B model with DFlash went from 11.98 to 41.75 tokens per second at about 7.5 W of power consumption. Energy per token fell from 0.627 to 0.182 J. Across four models the speedup was 2.21–3.49x, and in an automotive assistant test 1.35–2.06x, with 900 requests out of 900 handled correctly. Palette 3.0 allows several models and separate Docker containers to run in parallel on a single chip. The GPT-OSS-20B mixture-of-experts model is supported on the 16 GB Modalix module; the code is open on GitHub.