Google's Gemma 4 adds MTP for faster local inference, VibeVoice ported to C++, Ollama gets desktop layer

Google has released Gemma 4 with Multi-Token Prediction (MTP), a feature that allows the model to predict multiple tokens simultaneously, significantly speeding up local inference. Additionally, a C++ port of Microsoft's VibeVoice model, vibevoice.cpp, has been developed using the ggml library, enabling advanced speech-to-text and text-to-speech capabilities on consumer hardware without Python. A separate project is also underway to create an offline, low-RAM desktop application for Ollama, aiming to simplify local LLM deployment for less technical users. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Accelerates local LLM deployment and multimodal AI capabilities on consumer hardware.

RANK_REASON This cluster details updates to open-weight models and ports of existing models for local deployment, rather than a new frontier model release. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Google
• Gemma 4
• MTP
• Microsoft
• VibeVoice
• vibevoice.cpp
• ggml
• llama.cpp
• Ollama

• model release
• product
• infra