For companies putting AI into a robot, drone or industrial camera, a model's intelligence is only part of the problem. The device also has to process sensor inputs, generate useful responses and keep other applications running without exhausting its memory or power supply.
SiMa.ai is targeting that deployment problem with Palette 3.0, its on-chip software operating system that expands the AI workloads that can run on its Modalix edge AI processors. Palette 3.0 improves how applications share the chip and introduces techniques the company says can accelerate text generation by as much as 3.5X.
The platform targets developers building AI into physical products, including robots, drones and industrial systems. The Palette Neat core library, part of its application-development layer, carries an Apache 2.0 license. The license terms for the compiler and firmware remain unconfirmed.
Developers can start through SiMa.ai's Developer Center and GitHub organization. The company's DevKit 3.0 store listing prices a development system at $1,499 for the U.S. and Europe. It includes a Modalix module, carrier board, enclosure, power supply and 500GB of storage. The announcement does not provide a separate Palette 3.0 software price or production-volume hardware rates.
The Palette SDK has been available since October 31, according to a SiMa.ai community notice. The company says its upgraded Model Browser goes live November 17 as it details the update's capabilities and performance.
In an interview with VentureBeat, SiMa.ai GenAI product manager Manuel López Roldán confirmed that the changes run on existing silicon.
“It’s all software,” he said. Ashok Sudarsanam, the company’s associate vice president of machine learning accelerator (MLA) software, said his team builds the compiler that translates models into instructions for the accelerator and the software that schedules them to run.
“From day one, we built our hardware and software together,” Sudarsanam told VentureBeat. He said SiMa.ai updates the software to support newly released models and speed up applications on the chips customers already have.
Running several AI workloads on the same chip
Palette 3.0's central change concerns what happens when a device needs to do more than one AI task.
SiMa.ai’s media presentation shows one Modalix chip running a language model alongside three vision-language models. A separate demo runs six Docker containers sharing the chip’s resources across tasks ranging from object detection to 3D pose estimation.
The company changed how applications submit work to the chip. Processes now send requests directly to drivers in the operating system kernel, bypassing an intermediate service. Each hardware engine handles one request at a time, while different engines can do separate work in parallel. Developers can assign priorities to individual requests.
López Roldán gave VentureBeat the example of a camera feed that needs continuous object detection while a language model adds occasional context about the scene. He said developers can give the detector’s requests higher priority and let the language model use the accelerator as it becomes available. The demos illustrate how the chip shares resources. The briefing materials leave open how long high-priority requests might wait under sustained load.
SiMa.ai also demonstrated YOLOv26 object detection on 60 simultaneous 720p streams on one Modalix chip at 600 frames per second in aggregate, or an average of 10 frames per second per stream. In a separate camera-tuning workflow, López Roldán said developers can point a new camera at a color chart and use an AI agent to adjust image-processing settings. SiMa.ai says developers can make those adjustments without recompiling application code, which gives teams a way to tune a new sensor themselves.
What the 3.5x performance claim measures
SiMa.ai attributes its largest generation-speed gains to speculative decoding. The technique drafts several tokens ahead and lets the target model check them together, reducing the number of sequential steps needed to produce an answer.
The company's briefing presentation reports decoding-throughput improvements ranging from 2.21x to 3.49x relative to conventional generation on the same target models. The techniques include EAGLE3, DFlash and multi-token prediction for Gemma 4. One highlighted result uses DFlash N16 with Qwen3.5 9B to reach the upper end of that range.
The company measured those gains on Modalix against conventional decoding on the same models. In a separate 900-request automotive-assistant test, SiMa.ai reported a smaller speedup of 1.35x to 2.06x in response time. The company says all 900 requests produced the expected structured function calls.
SiMa.ai also estimates that the chip used 25.9% to 51.4% less energy per compute request while drawing roughly the same average power. In those tests, the chip saved energy by finishing the work sooner. The estimate covers the chip’s compute work. A robot’s total energy use also depends on components such as its sensors and motors.
Support for OpenAI's GPT-OSS-20B on a 16GB Modalix module is another highlighted capability. Fitting a model on the device is a useful milestone, but customers still need to measure its response time, context limits and accuracy under their intended configuration.
A model browser that shows how SiMa.ai got its numbers
López Roldán showed VentureBeat how the upgraded Model Browser exposes the trade-offs behind a benchmark. Along with throughput and latency, it shows measurements such as chip power, CPU and memory use, and how individual model layers execute. For models with multiple configurations, developers can see how lower numerical precision affects speed and accuracy. That can help a team choose which configuration to try after fine-tuning a model on its own data.
López Roldán said the browser’s download button provides an artifact prepared to run locally on Modalix. Developers who want to rebuild it can use scripts in a public repository that start with the original model from sources such as GitHub, Hugging Face or Roboflow. When it’s time to prepare a fine-tuned model for deployment, he said, “you can replicate the exact workflow.” He said SiMa.ai prioritized models its customers asked for when selecting the initial set.
“The board doesn’t even need an internet connection,” López Roldán said. He said customers can download update packages at a connected site and install them later on devices in the field. That matters for installations where a device needs to keep running with intermittent or absent connectivity.
How Palette compares with Nvidia and Hailo
SiMa.ai competes with a range of edge-computing options rather than a single interchangeable product. Nvidia supplies integrated Jetson development systems and its JetPack software environment. Hailo supplies accelerators that can add local AI processing to a host computer.
The listed prices cover development hardware. A deployment budget also needs to account for production hardware and integration.
Platform | Access and published hardware price | Software and deployment model | Relevant distinction |
|---|---|---|---|
SiMa.ai Palette 3.0 / Modalix | DevKit 3.0: $1,499; production quotes through SiMa.ai | Vendor hardware with an Apache-2.0 Neat core and additional platform components | Focus on multiple models sharing resources under a limited power budget |
Nvidia Jetson Orin Nano Super | Jetson hardware and Nvidia's software stack | Lower-priced entry into the Jetson ecosystem; not the same configuration as Modalix | |
Nvidia Jetson AGX Thor | Jetson hardware with JetPack; commercial platform rather than an entirely open-source system | Targets demanding robotics and physical-AI applications | |
Raspberry Pi AI HAT+ 2 with Hailo-10H | Add-on: $200, excluding the Raspberry Pi host | Accelerator add-on with Hailo software and model tooling | 8GB of dedicated memory for local language and vision-language workloads |
Nvidia's Jetson Thor technical overview describes connecting a development kit to an existing robot and using JetPack 7 to develop and test applications. The table uses Nvidia’s FAQ pricing.
The Hailo-based option offers a much cheaper hardware starting point, but the Raspberry Pi board, storage and other components must be added. Raspberry Pi's documentation describes support for language and vision-language models up to approximately 6 billion parameters. That differs from SiMa.ai's cited 20-billion-parameter mixture-of-experts example, although parameter counts alone do not establish accuracy or speed.
The enterprise decision is about the complete application
Palette 3.0 is most relevant to teams whose perception and reasoning models already work on their own but compete for compute and memory when they run together.
For a robot, that can come down to whether object detection responds within its deadline while the language model generates an answer.
SiMa.ai’s decoding figures put a number on generation speed. The mixed-workload demos leave that deadline question for developers to test on their own applications.
Existing customers also have migration work to account for. SiMa.ai’s release notice says moving from platform version 2.1.3 to 3.0.0 requires a migration, and developers must rebuild native applications against the matching components.
