Quick take

  • SiMa.ai announced Palette 3.0 for Modalix, allowing multiple models and Docker containers to share accelerator resources with adjustable priorities and parallel hardware engines.
  • SiMa.ai reported one Modalix chip running 60 simultaneous 720p object-detection streams at an aggregate 600 frames per second.
  • Its technical paper reported up to 3.5 times faster token generation, up to 2.06 times faster actionable responses and 51% lower compute energy per request.
  • Existing Modalix customers can adopt the update by recompiling applications, while VentureBeat reported a $1,499 developer kit and undisclosed production pricing.

Why it matters

For companies deploying robots, drones and industrial cameras, Palette 3.0 shifts edge-system design from assigning hardware to individual AI functions toward sharing one processor among several services, with existing Modalix installations able to gain those functions through recompilation.

A warehouse robot may soon watch a conveyor belt, understand a spoken instruction and plan its next movement on the same small computer, instead of sending each task to separate hardware or a remote cloud. SiMa.ai says a software update can bring that kind of multitasking to existing edge AI systems.

SiMa.ai announced Palette 3.0 for its Modalix edge AI processors, presenting the release as a way to make one chip behave less like a single-purpose accelerator and more like a shared workstation for physical machines. The company says the update can run language, vision and vision-language models concurrently, while improving text-generation throughput by as much as 3.5 times through speculative decoding. SiMa.ai said in its announcement that the software supports GPT-OSS-20B and is designed for systems such as robots, drones, industrial cameras and autonomous vehicles.

The practical appeal is easy to picture. A delivery robot could identify obstacles, estimate the position of a person, interpret a spoken request and generate a response without waiting for four separate computers to coordinate. A factory camera could inspect products while another application tracks workers, and a language model could remain available for operators asking questions about the production line.

Read LFM3 Brings Small, Multimodal AI Into the Enterprise’s Own Hands

That is the problem Palette 3.0 is intended to address. Edge devices often have strict limits on memory, energy and heat. They also need predictable response times. A model that performs well in a data center may be too expensive, too power-hungry or too slow when it must operate inside a drone or beside a moving machine.

One chip, several jobs

SiMa.ai says Palette 3.0 allows multiple models and Docker containers to share Modalix accelerator resources. Applications can receive different priority levels, while separate hardware engines work in parallel. The company demonstrated a language model running alongside three vision-language models, as well as six Docker containers handling tasks that included object detection and 3D pose estimation.

For developers, the change may be less about adding a dramatic new capability than about reducing the amount of hardware required to deliver an existing one. Instead of assigning a chip to each AI function, a system designer could divide the processor among several services and adjust their priority as conditions change.

The company’s community release notice says existing Modalix customers can adopt the update by recompiling their applications. SiMa.ai’s Palette SDK 3.0 notice describes concurrent language and vision processing, shared accelerator access and six-container execution. If that process works as intended, deployed systems could gain new functions without being physically replaced, an important consideration for fleets of robots or cameras installed across large facilities.

SiMa.ai also reported 60 simultaneous 720p object-detection streams on one Modalix chip, with an aggregate rate of 600 frames per second. That kind of density could matter in places where dozens of cameras must operate continuously but adding servers would increase cost, space and power consumption.

Still, a demonstration is not the same as a long-running deployment. The important questions include how performance changes when all workloads remain busy, how quickly lower-priority requests are served and whether a time-sensitive control system can tolerate unpredictable delays. A robot can accept a small delay when answering a question, but not necessarily when avoiding an obstacle.

The speed claim has conditions

The largest headline number comes from speculative decoding, a technique that uses a smaller or faster process to suggest several tokens before a larger model verifies them. When those predictions are accepted, text can be generated with fewer sequential steps.

In its technical paper, SiMa.ai reported up to 3.5 times faster token generation, along with up to 2.06 times faster actionable responses and 51% lower compute energy per request. The paper examines EAGLE3, DFlash and Gemma 4 MTP approaches on Modalix hardware.

Those figures describe chip computation rather than the complete energy footprint of a robot, drone or camera system. Motors, sensors, cooling, networking and other electronics can dominate a physical machine’s power use. Even so, lower accelerator energy could extend battery life or create room for more AI tasks within the same thermal budget.

The testing also shows why the number should be read carefully. SiMa.ai’s reported gains range from 2.21 to 3.49 times when compared with conventional decoding on the same models. The company separately described a 900-request automotive-assistant test with response-time improvements between 1.35 and 2.06 times.

Darius Popa of The Next Web wrote that the 3.5-times result came from SiMa.ai’s own testing on one model under favorable conditions, while the more product-like automotive test produced the smaller gains. That is the only outside assessment supplied for this story. No opposing company response or independent benchmark was provided, so there is not yet a second technical view to compare with SiMa.ai’s measurements.

From benchmark to daily use

Palette 3.0 also adds an upgraded Model Browser, which exposes measurements for throughput, latency, power, memory and individual model layers. That visibility could become as important as the raw speed improvements. Engineers building physical AI systems must decide whether a model should respond faster, consume less power or support more simultaneous streams. Those goals often conflict.

A developer might choose a smaller model for instant obstacle detection, reserve more memory for a language assistant and accept slower responses for a rarely used planning function. A tool that shows those trade-offs could make deployment decisions less dependent on guesswork.

VentureBeat reported that SiMa.ai is offering a $1,499 developer kit in the United States and Europe, while production pricing remains undisclosed. That leaves open the larger commercial question: whether software improvements can deliver enough value to justify adopting Modalix hardware at scale.

The broader idea is significant even if the benchmarks narrow in real-world use. Edge AI is moving toward machines that must perceive, reason and communicate at the same time, often without reliable connectivity. If software orchestration can make existing chips serve several roles, companies may be able to extend hardware lifetimes and deploy more intelligence locally.

The future promised by Palette 3.0 is not a robot with one spectacular model. It is a crowded little computer inside the robot, quietly dividing attention among many imperfect but useful systems. Whether that future arrives will depend on sustained performance, integration effort and timing under pressure, not only on the fastest result in a laboratory test.

Sources

Originally published at spaisee.com.

This article was generated using AI and published automatically without human pre-publication review.