SiMa.ai's Palette 3.0 Runs Several Physical AI Models on One Chip
SiMa.ai's Palette 3.0 software lets its Modalix MLSoC chip run multiple AI models at once in separate on-chip containers, promising up to 3.5x higher tokens per second within a low power budget, open source on GitHub, and aimed squarely at drone and robot integrators who today run one board per model.

Edge-AI compute is quietly splitting into two eras: one board per model, and one chip running several models at once. SiMa.ai said on November 17, 2026, in San Jose, California, that its new Palette 3.0 software suite lets a single Modalix MLSoC chip run multiple AI models concurrently, in separate on-chip Docker containers, within a low power budget, delivering up to 3.5 times higher model tokens per second (TPS) for robotics and drone workloads.
That matters to anyone specifying compute for a fleet of robots or drones, because the old assumption, one board per model, one power budget per board, one integration headache per vendor, is the assumption Palette 3.0 is built to break. For readers who have not tracked SiMa.ai closely, the company is a Physical AI chip and software maker headquartered in San Jose, California. Its core technology pairs low-power Modalix MLSoC silicon with the Palette software platform to run AI models at the physical edge, on drones, robots and cameras, putting it in direct competition with Nvidia's Jetson and Orin lines for the edge-robotics compute budget line item.
What Concurrent Model Execution Actually Changes for Integrators
Palette 3.0 builds on SiMa.ai's June 2026 "Agentic Palette" announcement, and the headline technical move is concurrency. Rather than dedicating a board, and its own power and thermal envelope, to each model a robot needs, Palette 3.0 lets a perception model, a depth model and a language or vision-language model (VLM) run as separate containers on the same Modalix MLSoC chip at the same time. SiMa.ai said the release folds in sensor-fusion tooling, a neural approach to depth estimation, and a closed-loop simulation workflow for digital twins, all built with unnamed ecosystem partners, with benchmarks now published in a new Model Browser.
For a robotics integrator today, running a large language model (LLM) alongside a vision pipeline on a drone or mobile robot typically means two boards, two power rails, two sets of thermal management, and two points of integration risk on an already weight- and power-constrained platform. Consolidating that workload onto one chip is not a convenience feature, it is a bill-of-materials decision. Fewer boards means fewer power supplies, less wiring harness, less enclosure volume, and fewer firmware images to validate and update across a fleet. On a drone, every gram and every watt not spent on a second compute board is a gram and a watt available for payload, sensors or flight time. On a mobile robot, it is one less thermal zone to manage and one less supply chain dependency to track across a production run.
The tradeoff integrators should not skip is concentration risk. Running everything on one chip means a single point of failure for every model a robot depends on, and it raises the stakes of thermal and power engineering on that one part. Buyers evaluating Palette 3.0 should ask SiMa.ai, or push through the vendor's competitive-benchmarking channel, for real sustained-throughput numbers under multi-model load, not just the headline 3.5x TPS figure, which SiMa.ai describes as achieved within a low power budget but has not published against a specific baseline configuration in its own materials.
Where This Sits Against Nvidia's Edge Silicon
Nvidia's Jetson and Orin families remain the default reference point for edge robotics compute, with a large existing developer base, broad framework support and years of integrator familiarity behind them. SiMa.ai is not trying to out-market that incumbency; it is trying to out-compete on power efficiency and on a software layer purpose-built for running several models concurrently rather than one model well. That is a narrower, sharper bet, and Palette 3.0 is the clearest evidence yet that SiMa.ai intends to compete on software capability, not just chip specifications, as the differentiator.
The practical question for a procurement manager is not which chip benchmarks higher on a single model, it is which platform's total cost of ownership is lower once you account for the number of boards, the integration engineering hours, and the power budget needed to run the actual multi-model stack a modern robot or drone requires. A chip that runs three models adequately on one power budget can beat a chip that runs one model superbly if the buyer's real workload is the former, not the latter. SiMa.ai is explicitly positioning Modalix MLSoC and Palette 3.0 to win that comparison, and the Model Browser benchmarks it published alongside the release exist specifically so buyers can run that math themselves rather than take the company's word for it.
Here is the quotable version of that argument: the real competitive question in edge-AI silicon is no longer how fast one chip runs one model, it is how many models a buyer can run per watt before the second board becomes unnecessary.
Open Source Code, Proprietary Silicon: What the GitHub Release Means for Lock-In
Palette 3.0's code is open source and available on GitHub, under the sima-neat organization, with documentation in SiMa.ai's Developer Center. That is a meaningfully different posture from the fully proprietary software stacks that typically ship with edge-AI silicon, and it is worth unpacking what it does and does not do for integrator lock-in risk.
Open-sourcing the software layer lowers one kind of lock-in: integrators can inspect, modify and extend the container orchestration, model-serving and sensor-fusion code themselves, rather than depending entirely on SiMa.ai's roadmap and support cycle to fix bugs or add features. That is a real advantage for integrators who have been burned by proprietary robotics middleware that stalled when a vendor's priorities shifted. It also means the code is auditable before it goes into a safety-relevant deployment, which matters more in robotics than in most software categories.
What open-sourcing Palette 3.0 does not do is remove the hardware dependency. The software still runs on Modalix MLSoC silicon, and the most advanced capabilities in this release, support for large Mixture-of-Experts (MoE) models such as GPT-OSS-20B, are tied specifically to the 16GB system-on-module (SoM) variant of that chip. An integrator adopting Palette 3.0 is still committing to SiMa.ai's silicon roadmap, pricing (undisclosed publicly, available on request through [email protected]) and supply chain, even while gaining more control over the software stack running on top of it. That is a genuinely better lock-in profile than a closed stack on closed silicon, but it is not the same as true platform independence, and buyers should price it accordingly rather than treat "open source" as a synonym for "no lock-in."
A Maturing Market, Not a New One
Palette 3.0 also adds generative AI features aimed squarely at making inference faster on constrained hardware: EAGLE3, DFlash, and Gemma 4 multi-token-prediction (MTP) speculative decoding support, all techniques for predicting multiple output tokens ahead of time to cut the effective latency of running language and vision-language models on-device. Paired with Mixture-of-Experts support on the 16GB SoM, the release signals that SiMa.ai is chasing the same generative-AI workloads data centers run, just compressed into a power envelope that fits inside a drone or a mobile robot.
That ambition sits alongside, and is distinct from, SiMa.ai's other recent news. The company raised US$150 million for physical AI chips in a funding round reported in late September 2026. Palette 3.0 is a separate software release, not a restatement of that funding news, and it is the clearer signal of where the money is being pointed: toward a software platform that can run increasingly sophisticated generative and perception models concurrently, on hardware that already exists in the field.
SiMa.ai founder and chief executive Krishna Rangasayee framed the release as delivering "hardware at the speed of software," adding that Palette 3.0 is "completely rewriting the roadmap for fully autonomous systems." That is a bold claim from a company still working to build name recognition against a much larger incumbent, and it should be read as a statement of intent rather than an independently verified industry shift. What it does accurately reflect is a company moving quickly from raising capital to shipping a differentiated software capability, which is the sequence buyers should want to see from any physical-AI chip vendor before committing a fleet to its platform.
SiMa.ai is running two webinars to walk integrators through the release directly: "Inside Palette 3.0: Any Physical AI Model. Real Machine Autonomy." on December 2, 2026, and "Drone and Robot Intelligence at Flight Speed: Build, Develop and Deploy Faster." on December 8, 2026. Both are a reasonable next step for a technical buyer who wants to see the concurrent multi-model execution claims demonstrated rather than described.
What Buyers Should Take From This
The shift from one-board-per-model to several-models-per-chip is a genuine inflection point for how robotics and drone platforms get architected, not a marketing reframe of incremental silicon progress. Concurrent execution changes the bill of materials, the power budget and the integration workload for any buyer running more than one AI model on an edge platform, which by 2026 describes most serious robotics deployments. Buyers should treat the specific numbers SiMa.ai has published, the 3.5x tokens-per-second figure most of all, as a starting point for their own benchmarking against their actual model mix, not as a settled fact to design around.
What is settled is the direction of travel. Edge-AI silicon vendors are now competing as much on software orchestration and multi-model efficiency as on raw chip performance, and open-sourcing that orchestration layer, while keeping the underlying silicon proprietary, is becoming a credible middle path between fully closed stacks and fully open hardware that does not yet exist at this performance tier. Integrators evaluating Modalix MLSoC and Palette 3.0 against Nvidia's Jetson and Orin lines should run their own concurrent-workload benchmarks before specifying either platform at fleet scale, because the winner of that comparison will be decided by actual throughput per watt on the buyer's own model stack, not by either vendor's press materials.
This article is for general information purposes only and does not constitute investment, financial, legal, or professional advice. It synthesizes company statements and public market activity as of publication.











