SiMa.ai launches Palette 3.0 to run several AI models on one chip

SiMa.ai announced on Tuesday, November 17, Palette 3.0, a new version of its software that lets several artificial intelligence models run at the same time on a single Modalix MLSoC chip, aimed at drones, robots and devices that need AI on the hardware itself, without relying on data centers.

According to the company, based in San Jose, California, Palette 3.0 builds on the Agentic Palette introduced in June 2026 and is designed to scale performance across any “Physical AI” model — the AI that lets machines reason and act in the physical world.

Several models on one chip

Palette 3.0’s headline feature is optimized concurrent execution: perception, depth, language and reasoning models now run in parallel on the same chip, in separate containers, with load balancing across multiple MLSoCs on a single board. For multimodel applications — such as running a vision-language-action (VLA) model alongside visual reasoning — the company promises up to 10x higher end-to-end throughput.

The announced numbers

The performance gains disclosed by SiMa.ai include:

  • Up to 3.5x more tokens per second (TPS) for Qwen2.5-VL-7B with EAGLE3 speculative decoding;
  • Up to 2.8x more TPS for Qwen2.5-VL-7B with DFlash;
  • Up to 1.8x more TPS for Qwen3-8B with 4-bit Gemma quantization;
  • Up to 60% lower time to first token (TTFT) for Qwen2.5-VL-7B;
  • Up to 2.3x more TPS for Depth Anything V2;
  • Up to 35% more frames per second (FPS) for RT-DETR and up to 28% for D-FINE.

Palette 3.0 also adds support for large Mixture-of-Experts (MoE) models, such as OpenAI’s GPT-OSS-120B, with up to 128 experts and up to 8 active experts per token.

Open source and Model Browser

Part of Palette 3.0’s code is open source and available on GitHub, under SiMa.ai’s organization, with documentation in the company’s Developer Center. The new Model Browser, at sima.ai/model-browser, gathers benchmarks for the key models behind fully autonomous drones and robotics, with results measured within real-world power limits.

The Palette 3.0 SDK had already been available since October 31, according to a company community notice; the November 17 announcement details the performance numbers and the new Model Browser. SiMa.ai has also scheduled two webinars about the release, on December 2 and December 8, 2026.

The numbers come from the company itself

One caveat: all the performance figures disclosed come from SiMa.ai’s own tests, with no independent verification so far. Founder and CEO Krishna Rangasayee said the bottleneck in Physical AI today is the mismatch between software, which evolves fast, and hardware, which takes years to design — and that the answer is a software-centric platform.

Image credit: SiMa.ai / BusinessWire.

Sources

Transparency note: this article was produced with the assistance of artificial intelligence, cross-checked against public posts and internet sources. Please consult the original sources for more details.

By Muse