SiMa.ai Speeds Language Models 3.5x With Palette 3.0
The new software update for Modalix AI chips raises language model text generation speed while keeping chip power draw broadly unchanged.
Updated on Nov. 17, 2026 in Tech

SiMa.ai has released Palette 3.0, the next version of software for its Modalix AI chips, which it says speeds up language models by up to 3.5 times. Power draw on the chip stays broadly the same.
Why it matters
Faster text generation on a chip with unchanged power draw could make running language models on local hardware more practical for developers. The gains come from speculative decoding, a method that lets a small draft model propose tokens for the main model to check.
In the in-car assistant test, all 900 measured requests returned the correct function call, and speculative decoding made them 1.35 to 2.06 times faster. The peak 3.5x figure rests on a single SiMa.ai white paper test, and results outside the vendor's own benchmarks are not yet reported.
The players
SiMa.ai
A San Jose-based company that develops AI chips and software for edge devices, with its Modalix line now supported by the Palette software platform.
The details
Palette 3.0 uses three speculative decoding methods, EAGLE3, DFlash and Gemma 4 multi-token prediction, and can run several models directly on the chip's accelerator with no background service in between. It can also run separate Docker containers on the chip and supports large mixture-of-experts models on the 16GB Modalix module.
Timeline
December 2025: Palette 2.0 added generative AI support on Modalix.
June: Palette Neat added an AI agent for building applications.
September: SiMa.ai raised $150 million at a $1.45 billion valuation.
November 17, 2026: Palette 3.0 was released.
The Tech Race
Palette 3.0 applies speculative decoding, a technique typically associated with large data-center inference, across three named methods on edge hardware. That moves a draft-and-verify approach onto chips with a 16GB memory ceiling, a comparison point for rival edge-AI vendors.
Developers building AI features on edge devices could run text-generation models faster without a higher power budget, if the speed-ups hold outside SiMa.ai's tests. Teams evaluating Modalix hardware can compare the reported tokens-per-second figures against their own models before committing to the Palette 3.0 workflow.
The takeaway
Speculative decoding is moving from large data-center systems onto smaller edge chips, and the speed gains reported here are strongest on the one model SiMa.ai tested directly. Readers should treat the 3.5x figure as a best case until independent benchmarks confirm the wider 2.21 to 3.49x range.
Further reading
For more on the broader field, see Artificial Intelligence coverage from San Jose.






