SiMa.ai's Palette 3.0 promises 3.5 times faster AI. Its own tests show where that number comes from

4 Sources

Share

SiMa.ai released Palette 3.0 on November 17, a software update for its Modalix chips that lets drones, robots and other machines run several AI models at the same time on one low-power chip

3

2

. The company says text generation on those chips can run up to 3.5 times faster

1

.

Source: The Next Web

Source: The Next Web

SiMa.ai is a San Jose startup founded in 2018 that builds chips and software for what it calls physical AI, meaning intelligence that runs on the device itself rather than in a data center, and it raised $150 million at a $1.45 billion valuation in September

3

1

. Founder and CEO Krishna Rangasayee framed the release as breaking a cycle in which software improves monthly and silicon takes years to catch up: "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."

4

The fine print on 3.5x

The speed claim rests on speculative decoding, a method where the chip drafts several likely next words at once and checks them in one pass. It speeds the generation of new text, not the processing of the prompt that came before, which is why the gains shrink on realistic workloads

1

2

.

The 3.5x peak comes from a single model under favourable conditions in SiMa.ai's own white paper

1

2

. Across four models the measured range was 2.21x to 3.49x, and a more realistic 900-request test showed 1.35x to 2.06x

1

2

. That test asked the chip to act as an automotive assistant, turning spoken commands such as a passenger saying they are cold into one correct function call

1

2

. Every request came back correct

1

2

.

Many models, one chip

The central change is concurrency. Multiple models, including separate Docker containers, run at once on a single Modalix chip while keeping its low power profile

3

2

. Requests now go straight to the kernel drivers with per-request priorities

3

2

. The company says the goal is to clear bottlenecks in integration and validation that hold up autonomous development, so perception and reasoning models can share one low-power chip

3

4

. In one demo, a single chip ran a language model alongside three vision-language models and six containers

2

.

Source: SiliconANGLE

Source: SiliconANGLE

"It's all software," said Manuel López Roldán, the executive in charge of agentic software and generative AI work

2

. Customers with chips already deployed get the update without buying new hardware

2

. The 16GB Modalix module can now run a 20-billion-parameter model from OpenAI

3

1

2

. A new Model Browser, live since November 17, publishes the company's measured throughput, latency, power draw and precision trade-offs

3

2

. The software can be downloaded to run offline, with update packages for devices already in the field

3

2

. Partner additions cover sensor fusion, depth sensing and digital twin simulation

3

4

.

The market, the price and the timing question

The company positions itself against incumbents it says are adapting data-center technology, while it builds silicon and software only for machines that act in the physical world

3

4

. Its named rivals include Nvidia's Jetson systems and Hailo's accelerators. Hailo's route is cheaper but handles smaller models, about 6 billion parameters against the 20 billion SiMa.ai demonstrated

1

2

.

The DevKit 3.0 development system costs $1,499 in the US and Europe; production-volume pricing has not been disclosed

2

. Existing customers have work to do: moving from version 2.1.3 requires a migration, and native applications must be rebuilt against matching components

2

. The core Palette Neat library is open source under an Apache 2.0 license, but how the compiler and firmware are licensed is still unconfirmed

3

2

.

The demos leave timing untested. Whether object detection keeps answering within its deadline while a language model generates on the same chip is something developers will have to test on their own applications

2

. The company's own briefing materials do not say how long a high-priority request might wait under sustained load

2

.

Source: VentureBeat

Source: VentureBeat

© 2026 TheOutpost.AI All rights reserved