The update runs several models at once on a single chip and adds open-source speculative decoding, with the company saying performance stays within a low power budget.
SiMa.ai, a company that makes chips and software for physical AI, such as drones and robots, has released Palette 3.0, the latest version of its software suite. It builds on the Agentic Palette that SiMa.ai announced in June, which the company describes as an industry first. SiMa.ai says Palette 3.0 improves how several AI models run at the same time, and raises the tokens per second (TPS) a model produces by up to 3.5 times while staying within a low power budget.
The release centres on SiMa.ai’s Modalix machine-learning system on a chip (MLSoC). According to the company, Palette 3.0 lets reasoning and multi-model workloads run concurrently on a single Modalix chip, including by running separate Docker containers on the chip. That is meant to support large language models and vision-language models within tight power and memory limits. SiMa.ai says the platform also speeds up the perception, depth and reasoning models that drones and robots rely on, and that, together with unnamed ecosystem partners, it adds unified sensor fusion, dense neural depth fields and closed-loop digital twin simulation to cut integration and testing delays in autonomous development.
The speed gain comes from speculative decoding, a technique in which a smaller model drafts text that a larger model then checks, so the large model produces tokens faster. Palette 3.0 adds the EAGLE3, DFlash and Gemma 4 MTP algorithms as open-source capabilities for building generative AI applications. SiMa.ai also says it now supports running large mixture-of-experts models, such as OpenAI’s GPT-OSS-20B, on its Modalix MLSoC 16GB system-on-module. A new Model Browser on its developer site shows how the key models perform within real-world power limits.
Krishna Rangasayee, founder and chief executive of SiMa.ai, said: “The fundamental bottleneck in Physical AI today is the glaring mismatch between software that evolves monthly and legacy silicon that takes years to adapt. We are breaking that cycle.” He said Palette 3.0 delivers “hardware at the speed of software”.
SiMa.ai says that, unlike incumbents that adapt data-centre technology, it builds software and silicon only for the demands of physical AI, and that its open-source code can be freely adapted. It also says its agentic software already cuts development time from months to days and hours.

No responses yet