Every vendor benchmark is a story about two numbers: the one on the slide and the one in the fine print. SiMa.ai's Palette 3.0 launch, announced in its own press release on November 17, 2026, is unusually honest about both. The headline says language models on its Modalix edge chips now generate text up to 3.5x faster. The second test, buried in the briefing materials covered by VentureBeat, reports speedups of only 1.35x to 2.06x on a realistic in-car assistant workload. Both numbers are real. The distance between them is where the actual engineering lives, and it is the number any team buying edge AI hardware should ask for.
Context: What Palette 3.0 Actually Ships
SiMa.ai is a San Jose startup, founded in 2018 by CEO Krishna Rangasayee, that builds the Modalix MLSoC, a purpose-built system-on-chip for what the industry has started calling physical AI: robots, drones, industrial cameras, vehicle systems. Customers and partners include Bosch, Emerson and Synopsys. The company raised a $150 million Series C at a $1.45 billion valuation in September, led by Fidelity Management & Research and Amplify Partners, bringing total funding to roughly $500 million, per SiliconANGLE.
Palette 3.0 is a software release, not a chip. SiMa.ai product manager Manuel López Roldán was direct with VentureBeat about it: "It's all software." The release has four substantive parts:
- Concurrent model execution. Multiple models (the demo runs one language model alongside three vision-language models, and separately six Docker containers) now share a single Modalix chip. Processes submit work directly to kernel drivers, bypassing an intermediate service, and developers can assign request priorities, so a safety-critical object detector can preempt a chatty language model.
- Speculative decoding using EAGLE3, DFlash and Gemma 4 multi-token prediction, the source of the 3.5x claim.
- Mixture-of-experts support, with OpenAI's GPT-OSS-20B fitting on the 16GB Modalix system-on-module.
- A Model Browser that publishes per-model benchmarks along with chip power, CPU and memory use, and layer-level execution data, plus ready-to-run artifacts. The core Palette Neat library ships under Apache 2.0, though compiler and firmware license terms remain unconfirmed.
Pricing: the DevKit 3.0 costs $1,499 in the US and Europe and includes a Modalix module, carrier board, enclosure and power supply. No separate software price was announced. One catch for existing customers: moving from platform 2.1.3 to 3.0.0 requires a migration and rebuilding native applications.
Where the 3.5x Comes From
The technique is speculative decoding, the same family of methods that dominates cloud inference cost discussions: a small draft model proposes several tokens ahead, and the full target model verifies them in one batched pass, collapsing many sequential decode steps into fewer. TNW's coverage gets the best-case number down to two digits: Alibaba's Qwen3.5 9B with DFlash went from 11.98 to 41.75 tokens per second, drawing about 7.5W either way. Across four models the speedup ranged from 2.21x to 3.49x.
Two details make that test legible, and credit to SiMa for publishing them. First, it ran on five fixed prompts, including a maths proof and a Python script, conditions the company itself describes as favorable. Second, the power figure stayed flat at roughly 7.5W, which means the energy-per-token improvement (0.627 to 0.182 joules) is arithmetic, not a power breakthrough.
The second test is the one that looks like a product. An automotive assistant had to convert natural-language commands such as "I'm cold" into exactly one correct structured function call, across 900 measured requests. All 900 returned the right call, and speculative decoding cut response times by only 1.35x to 2.06x. SiMa's own explanation, per TNW: the method does not speed up prompt processing.
Why Two Numbers, and What the Gap Tells You
Here is the analysis section this story deserves, because the gap between 3.49x and 1.35x is not marketing spin, it is Amdahl's law wearing a hat, and it applies to every deployment of speculative decoding, on edge chips and in data centers alike.
A request has two phases. Prefill processes the whole prompt in one parallel pass; decode emits output tokens one at a time. Speculative decoding accelerates only the second phase. So end-to-end speedup is bounded by the fraction of time spent decoding:
$$S_{e2e} = \frac{1}{(1-f_{decode}) + f_{decode}/S_{decode}}$$
Run that against SiMa's own two tests. The five-prompt generation test is nearly pure decode: short prompts, long outputs, so $f_{decode}$ approaches 1 and the measured end-to-end speedup approaches the decode speedup, 3.49x. The automotive test is nearly the opposite: a long system prompt full of tool definitions, then a tiny output of one function call. Take the reported 1.35x floor, assume the underlying decode speedup is near the top of the measured range (3.49x), and solve the equation for the decode fraction: $f_{decode} \approx 0.36$. At the 2.06x end the same math gives about $0.72$. Either way, a third to two thirds of the assistant's latency sits in prompt processing, which no amount of draft-model cleverness touches.
Sanity-check the energy claim the same way. At 11.98 tokens per second and 0.627 J/token, the implied power draw is 7.51W. At 41.75 tokens per second and 0.182 J/token, it is 7.60W. The two published figures reconcile to within 1.2% against the "roughly 7.5W flat" statement. This is a small thing, but it is exactly the check most vendor decks make impossible: SiMa's numbers close their own loop, which raises the credibility of the rest.
The data-scientist read on the acceptance dynamics: speculative decoding gains scale with the draft model's acceptance rate, and structured function-call outputs are short and brittle, one rejected token flushes the rest of the draft. Long-form generation (proofs, code, prose) is where drafts get accepted deep into the tree. The 3.49x case (a maths proof and a Python script) and the 1.35x case (single structured calls) are, technically, the best and worst case for the same algorithm on the same silicon.
One more number worth modeling, with assumptions labeled: GPT-OSS-20B on a 16GB module implies weights must occupy well under 16GB. At roughly 0.5 bytes per parameter (the MXFP4-style quantization the model ships with, an inference from public model cards rather than SiMa's release), 20B parameters is about 10GB, leaving ~6GB for activations, KV cache and the OS. MoE helps latency here, not memory: every expert stays resident, but only a fraction participates per token, so decode is bounded by active parameters, not total ones. That is why a "20B" model can feel like a small model at the token level while still needing the big module's RAM. Our own cost-side treatment of this tradeoff runs deeper in Mistral Le Chonk Goes Open: Tokens Are the Story.
What This Means Against Nvidia and Hailo
The competitive framing in VentureBeat's comparison is worth keeping in front of you: the Jetson Orin Nano Super dev kit lists at $399, SiMa's DevKit 3.0 at $1,499, Jetson AGX Thor at $5,499, and a Raspberry Pi AI HAT+ 2 with Hailo-10H at $200 (host board excluded, and Raspberry Pi's docs cap supported models at roughly 6B parameters). SiMa is not winning on sticker price. It is selling the one thing the cheap boards struggle with: several models, including a 20B-class MoE, sharing one chip under a single power budget with priority scheduling, and reproducible benchmark artifacts you can diff against your own workload.
That pitch rhymes with an infrastructure bet we covered in Nvidia Opens Its Racks to Korean NPU Chips, where GPUs handle prefill and NPUs handle decode in the same rack. Both stories rest on the same observation this article's math section made: prefill and decode are different workloads, and the industry is splitting them apart, at data-center scale in one case and at 7.5 watts in the other.
Outlook
Three things to watch. First, SiMa offers competitive benchmarking details on request, so third-party numbers against Jetson are a matter of when, not if; the Model Browser's layer-level telemetry makes cherry-picking harder than usual, and a downloaded artifact that runs on your own board is a far stronger evaluation signal than any normalized tokens-per-dollar table. Second, the concurrent-execution scheduler is the underdog feature: the briefing leaves open how long a high-priority request waits under sustained load, and for a drone that question is a certification question, not a performance one. Expect the first real-world reviews to hammer exactly that gap, because "six containers share a chip" is a throughput demo, while "the detector answers within its deadline while six containers argue" is an engineering result. Third, watch the migration cost: a version 2.1.3 to 3.0.0 transition that requires rebuilding native applications is a quiet tax on every existing deployment, and it resets the "hardware at the speed of software" promise for current customers even as it sells nicely to new ones. If your workload looks like SiMa's own 900-request test (long prompts, short structured outputs), budget for the 1.35x, not the 3.5x, and ask any vendor, at any altitude, for both numbers. A company that publishes the unflattering one is a company you can actually evaluate against.