Google launched its Gemma 4 open models this spring, promising a new level of power and performance for local AI. The tech giant’s approach to edge AI is set to become even faster with the release of Multi-Token Prediction (MTP) drafters for Gemma. Google describes these experimental models as leveraging a form of speculative decoding to predict future tokens, significantly accelerating generation compared to traditional token-by-token processing.
The latest Gemma models are built on the same underlying technology that powers Google’s frontier Gemini AI. However, they are specifically tuned to run locally rather than on Google’s cloud infrastructure. While Gemini is optimized for Google’s custom TPU chips—operating in high-performance clusters with ultra-fast interconnects and memory—the Gemma 4 models are designed to function on consumer-grade hardware. A single high-power AI accelerator can run the largest Gemma 4 model at full precision, and quantization techniques enable it to operate on a consumer GPU.
Gemma empowers users to experiment with AI directly on their own devices, reducing reliance on cloud-based AI systems that require sharing data with third parties. In a notable shift, Google updated the license for Gemma 4 to Apache 2.0, a far more permissive open-source license compared to the custom terms used for previous Gemma releases. Despite these advancements, local AI models still face hardware limitations for many users—an issue MTP aims to address.
Related articles
UK Tax Authority Deploys AI to Combat Fraud with Human Oversight
Human staff members will still check the AI's findings.
AI-Generated Citations Flood Academic Papers, Raising Concerns Among Researchers
Last summer, Peter Degen's postdoctoral supervisor came to him with an unusual problem: One of his papers was being cited too much. Citations are the...
Anthropic’s Claude Code: No Long-Term Roadmap, Focus on Model Improvements and Developer Feedback
SAN FRANCISCO—Amid an ever-expanding array of surfaces, growing demand for tokens and compute, and a rapidly evolving user base, Anthropic doesn't hav...
xAI Unveils Grok Build: New AI Coding Agent in Early Beta for Elite Users
It's in early beta and only available to SuperGrok Heavy subscribers right now.
Honda Unveils Next-Gen Hybrid Accord and RDX Prototypes, Shifts EV Targets to Hybrids
Honda revealed prototypes of two new hybrid models, an Accord sedan and the Acura RDX SUV, during its annual business briefing this week, built on a p...
Musk vs. Altman Trial: Closing Arguments Reveal Legal Missteps and Evidence Mount
Today was closing arguments in the Musk v. Altman trial, and I almost feel bad writing about the unbelievable demolition derby I just witnessed. Steve...
Meta Ray-Ban Display Glasses Now Support Gesture-Based Messaging in WhatsApp, Messenger, and More
Meta is rolling out new features to its Meta Ray-Ban Display smart glasses, including bringing the ability to write messages just with hand gestures t...
Elon Musk’s 'Jackass' Trophy Takes Center Stage in OpenAI Trial
Yesterday, in Musk v. Altman, before the jurors came in, Sam Altman's team passed up what looked - from a distance - like a little league trophy. It w...