Neuphonic open sourced NeuDecide, a 43 MB model that maps speech plus tool definitions directly to tool calls with arguments, skipping transcription. It runs on a single CPU thread and ships as three ONNX graphs with a WASM browser demo, Python package, and Hugging Face weights. On SLURP's 10-tool task it scores 72.4% tool accuracy, roughly 3x the Parakeet + FunctionGemma baseline.
Underdog released a demo game called Woof showcasing its 32MB voice model that runs locally on edge devices like phones and computers. The model enables speech interaction without cloud processing, countering the trend of large models requiring massive compute. It targets users and developers who want private, low-resource voice AI.
Underdog released a demo game called Woof showcasing its 32MB voice model that runs locally on edge devices like phones and computers. The model enables speech interaction without cloud processing, countering the trend of large models requiring massive compute. It targets users and developers who want private, low-resource voice AI.
A client-side TTS demo running the Ming-3.0-tts-0.5B model entirely in the browser on CPU via WebAssembly, with no server, API key, or GPU. The pipeline uses a Qwen2-0.5B LLM decoder to produce conditioning vectors plus a diffusion transformer for audio. It lets developers run speech synthesis fully locally with zero network requests after load.
A client-side TTS demo running the Ming-3.0-tts-0.5B model entirely in the browser on CPU via WebAssembly, with no server, API key, or GPU. The pipeline uses a Qwen2-0.5B LLM decoder to produce conditioning vectors plus a diffusion transformer for audio. It lets developers run speech synthesis fully locally with zero network requests after load.
parakeet.wgsl runs NVIDIA's Parakeet TDT 0.6B V2 English transcription model entirely in the browser using raw WebGPU compute shaders and SIMD WebAssembly. It transcribes 1 hour of audio in 20 seconds on an Apple M5, offering fast, accurate speech-to-text without server dependencies. Aimed at developers needing on-device ASR for web apps.
parakeet.wgsl runs NVIDIA's Parakeet TDT 0.6B V2 English transcription model entirely in the browser using raw WebGPU compute shaders and SIMD WebAssembly. It transcribes 1 hour of audio in 20 seconds on an Apple M5, offering fast, accurate speech-to-text without server dependencies. Aimed at developers needing on-device ASR for web apps.
Cactus released Needle2, a 14MB agentic LLM for tool call, device use, and structured extraction on phones, wearables, smart homes, small robots, and microcontrollers. The model runs a full session in 28MB of RAM and achieves 500 tokens/sec on a Raspberry Pi 5.
Cactus released Needle2, a 14MB agentic LLM for tool call, device use, and structured extraction on phones, wearables, smart homes, small robots, and microcontrollers. The model runs a full session in 28MB of RAM and achieves 500 tokens/sec on a Raspberry Pi 5.
Cactus Compute launched Whistle, an ultra-small ASR model for budget phones, wearables, smart home devices, and microcontrollers. The 55M-parameter (36M active) model outperforms Whisper base on most benchmarks at a fraction of the file size. The company's stated goal is compressing intelligence onto small, under-served devices rather than chasing SOTA via scale.
Cactus Compute launched Whistle, an ultra-small ASR model for budget phones, wearables, smart home devices, and microcontrollers. The 55M-parameter (36M active) model outperforms Whisper base on most benchmarks at a fraction of the file size. The company's stated goal is compressing intelligence onto small, under-served devices rather than chasing SOTA via scale.