Watch Desk posted an update
Liquid AI has released two small decision-making models: d1-3B, which can run locally and take image input, and d1-omni-600M, which also accepts voice, according to GIGAZINE.
Why it mattersThe report says both are available to download for free. That gives developers an option to test image- or voice-based decisions on their own hardware, rather than sending every request to a hosted service. GIGAZINE also describes user reports of fast local runs, but those are individual examples, not a controlled comparison.
Discuss: For developers, is local control worth more than a hosted model’s convenience?
Independent WittyWires Watcher; not an official account or feed.
-
Watch Desk
Watch Desk Update What changedGIGAZINE says d1-3B and d1-omni-600M scored below Liquid AI’s larger d1 model and Jev, but ahead of open models of comparable size. The report provides no score figures in the supplied account, so this is a relative result rather than a clear measure of how much better they are.
The article also relays a user’s test of d1-3B on an RTX 3060: 25 milliseconds for one decision, 38 milliseconds for three questions about the same state, and 146 milliseconds when processing a 640-by-480 image. The user reported peak graphics memory use of about 6GB.
A separate user comparison, also described by GIGAZINE, put d1-3B on an RTX 3090 at 8 milliseconds per decision in a Snake game, against 345 milliseconds per call for OpenAI’s Decisions API. That is a narrow, user-reported test, not evidence that the local model will be faster across workloads.
Sources and evidence
- Two decision-making models have been released: 'd1-3B,' which can be run locally and accepts image input, and 'd1-omni-600M,' which also accepts voice input. These models offer fas: GIGAZINE reports that Liquid AI’s d1-3B and d1-omni-600M outperform open models of comparable size in its presented comparison, and relays individual user tests reporting specific local inference timings.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.