In short
- Google released DiffusionGemma, a free heavy-duty model that generates entire blocks of 256 tokens simultaneously through word diffusion – hitting more than 1,000 tokens per second on an NVIDIA H100, four times faster than conventional models.
- The DiffusionGemma recording module required for local lighting does not have a public runtime—not in mlx-lm, not in LM Studio—making it ineffective in most consumer setups today.
- On NVIDIA NIM, the version came pre-programmed to 8,192 tokens – well below the 64,000-token floor that agents like Hermes Agent require – meaning that autonomous processes cannot run without manual reconfiguration.
Google abandoned DiffusionGemma todayAn open-source AI that creates text the way photographers create images: start with noise, clean it up until it makes sense. It hits 1,000 tokens per second on the NVIDIA H100. (Symbols are an important part of the information that the AI model handles.) This means it’s four times faster than a regular Gemma. It’s free, Apache 2.0, with features on Hugging Face.
The grip, as always, is well printed. Per Google announcementexample hits “700+ tokens per minute on NVIDIA GeForce RTX 5090.” It also follows the standard Gemma 4 in output.
Google does it themselves. This is an example of speed, not quality improvement.
What this does
Every LLM you have applied is a typo. One sign at a time with each word dependent on the last. This is how autoregressive construction works.
DiffusionGemma does not do that. Instead of creating symbols sequentially, it starts with clean chunks of text mixed together. On Google design guide“starts with a screen of placeholder symbols” and repeatedly blocks the trusted symbols until the entire block is displayed. Two hundred and sixty tokens per front pass. The GPU is busy.
The result is double vision—each sign can see every other sign being created, which is not possible in independent models (they cannot see the future, which will be written). This makes it surprisingly good for tasks where the end of the solution precludes the beginning: code completion, optimized output, heavy workloads, and so on. Google has successfully edited the Bible to solve Sudoku as a presentation. The basic model got almost 0% of the images correct.
The improved version hit 80%.
Text publishing has been a research activity for many years. MDLM, A SEAT, LLaDA, Dreams– case studies that proved the method worked on a small scale and mostly serve as proof of concept. Inception Labs sent Mercury 2 in February 2026 as the first model of commercial distribution, which is said to be five times faster than the competition.
But nothing was heavy, and nothing came with zero-day support in vLLM, Hugging Face Transformers, and Unsloth. DiffusionGemma is the first open source release from the tier-one lab.
There are also historical jokes to consider. Image generators started as diffusion models (hence the name Stable Diffusion) and are now moving to stable architectures for better performance. Linguistics started out as random and now trying to spread rapidly.
Why does it hurt to run… right now
Running DiffusionGemma requires a marker – a lightweight component that provides parallel blocks, which the main type guarantees one step forward. This is called speculative decoding. DFlash is a framework published as early as 2026 that uses a small spread model as a model, supporting at 6x speed for other jobs. It is the engine that makes this model class useful.
Problem: DiffusionGemma needs a different model to run locally via MLX—Apple Silicon’s machine learning platform. That part is not available in any mlx-lm version, in the open pull request, or in the LM Studio runtime.
We tried running DiffusionGemma and Hermes via NVIDIA NIM. Example loaded, but then: “agent init failed: Model google/diffusiongemma-26b-a4b-has an information window of 8,192 symbols, which is below the 64,000 required by the Hermes Agent.”
To be fair: The actual DiffusionGemma window is 256K tokens. The number of 8,192 was Nvidia messing things up randomly, not the limits of its design.
In practice, configuring it properly for use requires manual work that everyday users are not yet familiar with, and Hermes Agent cannot start without it. The same speed means nothing if the agent can’t start.
Hopefully, in the next few days, the community will develop better tools to manage these models.
Who is it really?
Developers with NVIDIA RTX 4090 or 5090 hardware building materials for real-time online controllers, final, full code, stable generation. That’s the target. Like Decrypt written in MayGoogle has been forced to make local thinking faster without new tools.
For researchers, two-way sequencing opens up a field that independent models can’t reach—protein sequences, mathematical graphs, anything where N sites depend on N+50 sites. That is no small thing.
Google launched Gemma 4 under Apache 2.0 in April, and DiffusionGemma continues that path. There is already a llama.cpp PR open as of today. When the equipment arrives, it reaches more people.
On a machine with a reliable GPU, 1,000 tokens per second is realistic.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
