SiMa.ai details Palette 3.0, claiming up to 3.5× faster AI text generation
The software runs on existing Modalix chips. An upgraded Model Browser exposes performance tradeoffs, while application-level tests show smaller gains than the headline decoding figure.
Palette SDK 3.0 adds workload scheduling and model-inspection tools for SiMa.ai’s Modalix edge chips, aiming to run multiple AI tasks without replacing hardware. In November 17 disclosures, the company reported up to 3.49× decoding throughput with DFlash N16 and Qwen3.5 9B, but that figure measures text generation rather than full application response time. The software was released October 31, 2026; customers upgrading from version 2.1.3 must complete a full migration, not an in-place update.
01
SiMa.ai demonstrated one chip running a language model alongside three vision-language models, and another running six containers with vision tasks.
02
A 900-request automotive-assistant test reportedly cut chip energy per compute request by 25.9%–51.4% at roughly unchanged average power; whole-device energy was not measured.
03
The upgraded Model Browser exposes throughput, latency, chip power, CPU and memory use, and model-layer execution to help developers assess configuration tradeoffs.
SiMa.ai is promising more AI work from chips customers already own, but its biggest speed figure measures text generation—not a complete application’s response. In November 17 disclosures about Palette 3.0, the company reported up to 3.49 times higher decoding throughput and said an upgraded Model Browser would go live that day to expose performance tradeoffs.
The software itself arrived earlier. SiMa.ai’s November 2 community announcement dates Palette SDK 3.0’s release to October 31, 2026. The November 17 briefing adds detailed measurements and the browser upgrade around that release, rather than marking the software’s first availability. Product manager Manuel López Roldán told VentureBeat that the changes require no new silicon.
One chip, competing requests
Palette 3.0 targets robots, drones and industrial devices that need several AI tasks running locally. Independent processes and Docker containers—separately packaged applications—can share Modalix’s accelerator and video engines. SiMa.ai demonstrated one chip running a language model alongside three vision-language models, which process images and text. Another demonstration ran six containers performing tasks including object detection and 3D pose estimation.
The change sits in how applications submit work. Requests now go directly to drivers in the operating system’s kernel, bypassing an intermediate service. Each hardware engine handles one request at a time, while different engines can work in parallel. Developers can assign priorities to individual requests rather than treating all work equally.
López Roldán described giving continuous camera-based object detection priority while a language model uses spare accelerator time to explain the scene. That illustrates the intended balance between perception and reasoning. How long urgent requests might wait under sustained load remains unresolved; the mixed-workload demonstrations did not establish that deadline behavior.
Decoding speed is not total waiting time
The largest reported gains come from speculative decoding: drafting several pieces of text ahead, then having the target model check them together. This reduces sequential generation steps. SiMa.ai’s presentation compared that approach with conventional decoding on the same target models and Modalix hardware. Its highest result used DFlash N16 with Qwen3.5 9B.
In the automotive-assistant test, SiMa.ai said all 900 requests produced the expected structured function calls. It also estimated 25.9% to 51.4% less energy per compute request at roughly unchanged average power, because the chip finished sooner. That estimate covers chip computation, not the energy consumed by an entire robot’s sensors, motors and other components.
Inspect the configuration, then deploy
The upgraded Model Browser gives developers more than a headline speed score. It shows throughput, latency, chip power, CPU and memory use, and individual model-layer execution. Where multiple configurations exist, developers can inspect how lower numerical precision changes speed and accuracy before choosing a configuration to try with their own fine-tuned model.
López Roldán said its download button supplies an artifact prepared to run locally on Modalix. Public repository scripts let developers rebuild from original models and repeat the preparation workflow after fine-tuning. Devices can also operate without an internet connection: customers can download update packages elsewhere and install them later in the field.
Access and licensing
Development hardware: SiMa.ai lists DevKit 3.0 at $1,499 in the U.S. and Europe, including a Modalix module, carrier board, enclosure, power supply and 500GB of storage.
Software terms: the Palette Neat core library uses Apache 2.0. VentureBeat could not confirm the compiler and firmware license terms; no separate Palette 3.0 software price was disclosed.
Existing customers face a concrete upgrade task. Moving from platform 2.1.3 to 3.0.0 requires a full migration because the new version changes the operating-system and update layout. Native applications must be rebuilt against matching components, packages reinstalled and workloads validated; this is not a routine in-place update.
Reader comments
Newest comments first. Replies stay oldest first.