Google's DiffusionGemma Generates Text 4x Faster by Ditching Sequential Tokens

Google closed out the week with five simultaneous launches spanning consumer apps, developer tools, and open research. The announcements touch nearly every layer of the AI stack, from a new audio model that translates speech in real time to an experimental open model that throws out the standard text-generation playbook entirely.

Real-time voice translation, finally

Gemini 3.5 Live Translate is Google's new audio model built specifically for live speech-to-speech translation. Unlike turn-by-turn systems that wait for the speaker to finish before responding, the model generates speech continuously, balancing the trade-off between waiting for context to improve quality and translating immediately to stay in sync with the speaker. The result is a conversation that flows without the awkward pauses that plague most real-time translation tools.

The model automatically detects 70+ languages and generates smooth, natural-sounding translated speech that preserves the speakers' intonation, pacing, and pitch. All audio output is watermarked with SynthID, Google's imperceptible audio watermark, to keep AI-generated content detectable.

The rollout is happening across multiple surfaces at once:

• Developers: Available now in public preview via the Gemini Live API and Google AI Studio
• Enterprises: Private preview in Google Meet this month, with support for 70+ languages and 2,000+ language pair combinations, up from the previous limit of five languages and English-only pairs
• Consumers: Rolling out now in the Google Translate app on Android and iOS globally

For Android users, a new listening mode lets you hear translations directly through your phone's earpiece, simply by holding your phone to your ear like a regular call. Grab, the Southeast Asian ride-hailing platform whose users make over 10 million voice calls per month, is already testing the model for multilingual driver-passenger communication.

NotebookLM becomes an actual research agent

Google has officially overhauled NotebookLM with powerful agentic capabilities, advanced multi-step reasoning, and an isolated coding environment. The upgrade shifts NotebookLM from a passive reading assistant into an active research partner that writes code, discovers sources, and builds project repositories from scratch.

The new capabilities move the product from a document-analysis tool toward an autonomous research agent capable of writing code, browsing the web, and producing downloadable reports, though the new capabilities are limited to paid subscribers at launch. The underlying engine is now Gemini 3.5 and Antigravity, Google's specialized multi-agent coding framework introduced at I/O 2026.

The practical upgrades are substantial:

• Secure cloud computer per notebook: Every notebook now acts as an isolated virtual machine. NotebookLM can autonomously write and execute Python code in the background to handle heavy data calculations, mathematical modeling, and dataset normalization.
• Visible reasoning: You now have step-by-step visibility into the AI's logic directly in the chat interface, making answers completely transparent.
• Rich export formats: Users can now create downloadable outputs including PDF reports with charts and tables, Microsoft Excel spreadsheets, PowerPoint presentations, structured data files, images, and documents in formats such as DOCX, Markdown, CSV, JSON, PNG, JPG, GIF, SVG, and TXT.
• Live source discovery: NotebookLM can now help users build source repositories directly within chats, locate related materials, identify primary sources in different languages, and use Google Search to discover relevant information.

Google AI Ultra became substantially more accessible after a price cut at Google I/O 2026, and the NotebookLM update was the first major feature wave to land on that tier afterward, meaning Google made the agentic tier cheaper, then immediately gave it powerful new capabilities. The upgrade is web-only at launch.

Project Genie goes global

Google's Project Genie is an experimental AI world-model prototype that generates explorable 3D environments from text or image prompts. It was previously limited to US-based Google AI Ultra subscribers. This week, it expanded to Google AI Ultra subscribers worldwide.

Built on DeepMind's Genie 3 world-model research, the prototype lets users move through AI-generated scenes in real time and regenerate variations using revised prompts, rather than serving as a full game engine or production tool. Project Genie has been used for training AI agents in three-dimensional environments and video game design.

The global expansion matters for context: the move signals Google's aggressive push into the world model race as competitors like World Labs and Runway accelerate their own launches. Access requires a Google AI Ultra subscription and users must be 18 or older.

Gemini Notebooks expands to Europe

Notebooks in the Gemini App, Google's integrated notebook experience inside the main Gemini interface, are now available in the European Economic Area, the United Kingdom, and Switzerland. The expansion brings a feature that was previously geographically restricted to a large new user base, though no new functionality was announced alongside the rollout.

DiffusionGemma: a different kind of text generation

Google introduced DiffusionGemma, an experimental open model that explores text diffusion, an exceptionally fast approach to text generation. Released under an Apache 2.0 license, this 26B Mixture of Experts model moves beyond the sequential token-by-token processing of typical autoregressive LLMs.

To understand why this matters, it helps to know how standard LLMs work. Every model you've used, from GPT to Gemini to Claude, generates text one token at a time, left to right, each word depending on the previous one. DiffusionGemma does something different: it starts from a canvas of random tokens and refines them in parallel, a bit like image models do when they turn noise into a sharp photo.

Concretely, the model works on a 256-token block at a time. It starts with random placeholder tokens, then runs multiple denoising passes where confident tokens help resolve neighboring positions until the whole block snaps into coherent text. For sequences longer than 256 tokens, it commits each finished block to memory and starts a fresh canvas conditioned on what came before.

The speed numbers are striking. Built on the Gemma 4 backbone, DiffusionGemma delivers up to 4x faster token generation on GPUs, reaching 700+ tokens per second on an NVIDIA GeForce RTX 5090 and 1,000+ tokens per second on a single NVIDIA H100. It runs with just 18 GB of DRAM or VRAM.

The speed gain comes from a fundamental architectural shift. Standard LLMs are bottlenecked by memory bandwidth because they must reload model weights for every single token. DiffusionGemma shifts that bottleneck to compute instead, giving the GPU a large parallel workload that keeps its tensor cores busy rather than waiting on memory reads.

There is a real trade-off, though. On the same task, DiffusionGemma stays below standard Gemma 4 on every published benchmark. Google is transparent about this. The model is positioned for use cases where speed matters more than peak quality:

• Code infilling and inline editing
• Rapid iteration workflows
• Constrained generation tasks (the team released a fine-tuned Sudoku solver that hits 80% accuracy after a simple training recipe, versus near 0% for the base model)
• Local, single-user GPU deployments where latency is the primary constraint

DiffusionGemma uses bidirectional attention to evaluate the entire text block simultaneously during generation, enabling real-time error correction and parallel context propagation. This is a capability autoregressive models fundamentally cannot have: once a token is generated, it cannot be revised. DiffusionGemma can re-noise low-confidence tokens and try again within the same generation pass.

You can serve it today using vLLM with an OpenAI-compatible endpoint:

vllm serve google/diffusiongemma-26B-A4B-it \
  --max-model-len 262144 \
  --max-num-seqs 4 \
  --gpu-memory-utilization 0.85 \
  --attention-backend TRITON_ATTN \
  --generation-config vllm \
  --hf-overrides '{"diffusion_sampler": "entropy_bound", "diffusion_entropy_bound": 0.1}' \
  --diffusion-config '{"canvas_length": 256}' \
  --enable-chunked-prefill

Weights are available on Hugging Face under Apache 2.0. Fine-tuning recipes are available via the Hackable Diffusion JAX toolbox, and the model is supported by vLLM, Hugging Face Transformers, SGLang, and MLX out of the box.

Taken together, this week's five launches show Google pushing on every front simultaneously: production-ready audio AI for consumers and enterprises, a research tool that now competes with full agentic platforms, a world-model experiment going global, geographic expansion of existing features, and a research bet on a fundamentally different approach to text generation. The breadth is deliberate, and the pace is not slowing down.