Cactus Compute
A San Francisco-based AI startup incubated by Y Combinator, focused on developing ultra-compact on-device language models. Its primary product line is the Neural series of models, aimed at enabling efficient structured output and tool-calling capabilities on edge devices.
Timeline (last 90 days)
Cactus Compute launched the Whistle speech model, with a size of 16.9 MB.
The Whistle speech transcription model released by Cactus Compute is only 16.9MB (about 17MB) in file size.
Cactus Compute was previously known for its Needle series of models (tool-calling and structured-output models).
Cactus officially states that Needle is a task-specific model that should be fine-tuned before going to production, and that the base model is a starting point rather than a finished assistant.
Neural 3 (also known as Needle) is an on-device model launched by San Francisco startup Cactus Compute, incubated by YC; it requires no internet connection and can run on phones, watches, and small development boards.
Cactus replaces the feed-forward layer of the traditional Transformer with a headband MLP, containing only approximately 25,600 parameters per layer instead of 4.7 million, with most knowledge stored in an engram table that retrieves short 2-to-3-word fragments via hash lookup.
Cactus Compute's Niddle ranked second on the GitHub trending list, gaining 3,409 new stars.
All Facts (6)
Cactus Compute launched the Whistle speech model, with a size of 16.9 MB.
50%UnverifiedThe Whistle speech transcription model released by Cactus Compute is only 16.9MB (about 17MB) in file size.
50%UnverifiedCactus Compute was previously known for its Needle series of models (tool-calling and structured-output models).
50%UnverifiedNeural 3 (also known as Needle) is an on-device model launched by San Francisco startup Cactus Compute, incubated by YC; it requires no internet connection and can run on phones, watches, and small development boards.
50%UnverifiedCactus replaces the feed-forward layer of the traditional Transformer with a headband MLP, containing only approximately 25,600 parameters per layer instead of 4.7 million, with most knowledge stored in an engram table that retrieves short 2-to-3-word fragments via hash lookup.
50%UnverifiedCactus officially states that Needle is a task-specific model that should be fine-tuned before going to production, and that the base model is a starting point rather than a finished assistant.
50%