<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>Palette 3.0 &#8211; Tech Raman</title>
	<atom:link href="https://techraman.com/tag/palette-3-0/feed/" rel="self" type="application/rss+xml" />
	<link>https://techraman.com</link>
	<description>Simplified Tech Talk</description>
	<lastBuildDate>Thu, 19 Nov 2026 14:33:37 +0000</lastBuildDate>
	<language>en-US</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=7.1.3</generator>

<image>
	<url>https://techraman.com/wp-content/uploads/2020/01/TR_logo-150x120.png</url>
	<title>Palette 3.0 &#8211; Tech Raman</title>
	<link>https://techraman.com</link>
	<width>32</width>
	<height>32</height>
</image> 
	<item>
		<title>SiMa.ai Palette 3.0 targets robot AI workloads: what the 3.5x claim means</title>
		<link>https://techraman.com/sima-ai-palette-3-0-robot-ai-speed-claim/</link>
					<comments>https://techraman.com/sima-ai-palette-3-0-robot-ai-speed-claim/#respond</comments>
		
		<dc:creator><![CDATA[TechRaman News Desk]]></dc:creator>
		<pubDate>Thu, 19 Nov 2026 14:33:37 +0000</pubDate>
				<category><![CDATA[News]]></category>
		<category><![CDATA[edge AI]]></category>
		<category><![CDATA[Palette 3.0]]></category>
		<category><![CDATA[Physical AI]]></category>
		<category><![CDATA[Robotics]]></category>
		<category><![CDATA[SiMa.ai]]></category>
		<guid isPermaLink="false">https://techraman.com/?p=20842</guid>

					<description><![CDATA[A drone or robot often needs several AI models working together: one to understand a camera feed, another to estimate depth, and another to turn an instruction into an action. SiMa.ai&#8217;s Palette 3.0 software update...]]></description>
										<content:encoded><![CDATA[<p>A drone or robot often needs several AI models working together: one to understand a camera feed, another to estimate depth, and another to turn an instruction into an action. SiMa.ai&#8217;s Palette 3.0 software update targets that problem on its Modalix chips, with better concurrent model execution and faster text generation within a limited power budget.</p>
<p>The India press release is dated November 19, 2026. SiMa.ai&#8217;s global announcement was published on November 17, while its developer-community release note says the SDK became available on October 31. This is therefore an update to an existing developer platform, not a new chip launching today.</p>
<p><em>Image: Official SiMa.ai Palette 3.0 promotional artwork, supplied with the press release.</em></p>
<h2>Several models, one Modalix chip</h2>
<p>Palette 3.0 improves how multiple workloads share a single Modalix MLSoC, including applications running in separate Docker containers. The company lists large language models, vision-language models, perception, depth and reasoning workloads among its targets.</p>
<p>For a robotics team, the practical point is being able to combine those jobs on an embedded system rather than judge the chip by one isolated model&#8217;s speed. SiMa.ai also describes partner capabilities for sensor fusion, depth estimation and digital-twin simulation. These are development tools and supported workflows, not evidence that a robot becomes fully autonomous simply by installing the update.</p>
<p>The release&#8217;s broad claim about supporting any Physical AI model still needs to be checked against the actual model, memory footprint, operators and deployment configuration. A supported demonstration does not establish universal compatibility.</p>
<h2>What the up-to-3.5x speed claim measures</h2>
<p>The headline performance gain comes from speculative decoding. Instead of making a large target model produce every token one at a time, a lighter drafting mechanism proposes several candidates and the target model checks them together. When enough candidates are accepted, the system can generate an answer with fewer sequential verification steps.</p>
<p>SiMa.ai lists EAGLE3, DFlash and Gemma 4 multi-token prediction implementations in its LLiMa inference framework. In the company&#8217;s published example, DFlash N16 with Qwen3.5 9B increased decoding throughput from 11.98 to 41.75 tokens per second, about 3.49 times, at approximately 7.5W mean SoC power.</p>
<p>That is a vendor-reported comparison with conventional decoding on the same model and platform. It is not a claim that every model, complete application or competing chip is 3.5 times slower. We have not independently reproduced the benchmark.</p>
<p>A separate company automotive-assistant test illustrates the difference between throughput and useful response time. With Gemma 4 E4B and multi-token prediction, a complete structured tool call took 0.569 seconds rather than 1.170 seconds, about a 2.06-times improvement. SoC power is also not total robot power: cameras, other electronics and motors contribute to the full system&#8217;s consumption.</p>
<h2>MoE support and developer migration</h2>
<p>The release adds support for mixture-of-experts models such as OpenAI&#8217;s GPT-OSS-20B on a 16GB Modalix system-on-module. That shows a supported configuration, not that every 20-billion-parameter model fits or runs at the same speed. Developers still need to check the model build and runtime requirements.</p>
<p>The SDK release note says Neat 0.6.0 requires Platform 3.0.0, with native applications rebuilt against matching headers and PyNeat reinstalled. Moving from Platform 2.1.3 to 3.0.0 is not a regular update because the platform moves to Debian 13 and A/B system images. Existing teams should follow the migration instructions and validate their own workloads rather than treat it as a drop-in upgrade.</p>
<p>SiMa.ai provides code through its <a href="https://github.com/sima-neat" target="_blank" rel="noopener">Palette GitHub organisation</a> and a <a href="https://developer.sima.ai/models" target="_blank" rel="noopener">Model Browser</a> for model information. Its <a href="https://sima.ai/blog/breaking-the-single-token-bottleneck-with-speculative-decoding-on-modalix/" target="_blank" rel="noopener">speculative-decoding benchmark explanation</a> gives the context behind the speed claim.</p>
<p>For developers building drones, robots or other on-device AI products, the useful advance is more work from existing Modalix hardware. The next test is whether that gain holds when perception, reasoning and control run together under the product&#8217;s actual latency, memory, thermal and power limits.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://techraman.com/sima-ai-palette-3-0-robot-ai-speed-claim/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
			</item>
	</channel>
</rss>
