SiMa.ai details Palette 3.0 for concurrent edge AI workloads
Palette 3.0 adds concurrent model execution, shared accelerator access for containers, mixture-of-experts support and generative AI tools for SiMa.ai's Modalix hardware.
SiMa.ai detailed Palette 3.0 on November 17, adding concurrent execution for language, vision-language and other AI models on its Modalix edge AI hardware. The software lets multiple workloads share the chip’s accelerator and video engines within the same power and memory limits.
The software is intended to consolidate workloads that might otherwise run separately. A SiMa.ai-administered release notice says independent processes and Docker containers can share the hardware; the company’s example runs six applications side by side.
That notice records Palette SDK 3.0 as officially available on October 31, before the November 17 announcement. SiMa.ai’s current installation documentation also lists a release-3.0 SDK channel.
Palette 3.0 separates kernels, the Runner, Scheduler and Codecs into dedicated components with direct runtime and driver integration, which SiMa.ai says improves memory sharing. The release also adds mixture-of-experts support. SiMa.ai says GPT-OSS-20B has run on a Modalix system with 16GB of memory, but the available evidence does not establish its response time, context limits, accuracy or sustained-load behavior on that configuration.
The SDK updates SiMa.ai’s Neat software to version 0.6.0 while retaining its core C++ and Python Model, Graph and Run APIs. Native applications must be rebuilt, and moving a board from platform version 2.1.3 to 3.0.0 requires a documented migration rather than a standard update.
For generative AI, Palette 3.0 adds speculative-decoding options including EAGLE3, DFlash and Gemma 4 multi-token prediction. Speculative decoding uses a faster draft process to propose tokens that the main model then verifies. SiMa.ai claims these options can increase decoding throughput by up to 3.5 times, but the supplied evidence contains no independent test of that benchmark.
The company also introduced an updated Model Browser that displays throughput, latency, power, CPU and memory measurements, along with per-layer execution details for supported models.
More news

Meituan releases LongCat-2.5 weights under MIT license

Amazon researchers introduce HOPE expert-pruning method

Radical Numerics opens restricted Omnii-1 biodefense API and laptop deployment
