News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Gemini Diffusion

Gemini Diffusion

Experimental text diffusion model from Google DeepMind

Gemini Diffusion is an experimental text diffusion model from Google DeepMind. Instead of writing one token at a time, it refines noise to generate blocks of tokens quickly and correct errors as it goes, aiming at speed and control in text, maths and code generation.

Share on WhatsApp

  • Uses diffusion: refines noise step by step instead of predicting token by token
  • Generates whole blocks of tokens at once and can correct errors mid-generation
  • Sampling speed of 1479 tokens per second excluding overhead, 0.84 sec overhead
  • LiveCodeBench (v6) 30.9% versus 28.5% for Gemini 2.0 Flash-Lite, pass@1
  • Available as an experimental demo to help develop future models

What it is

Gemini Diffusion is an experimental text model from Google DeepMind that uses diffusion rather than the usual one-token-at-a-time approach. Diffusion models learn to build an output by repeatedly refining noise, which lets them revise their own work mid-generation. Google DeepMind presents it as a way to gain speed, control and creativity in text generation, and says it is especially good at editing tasks, including maths and code. The company reports sampling at 1479 tokens per second excluding overhead, with 0.84 seconds of overhead, averaged across the evaluations it reports.

How it came about

Google DeepMind describes Gemini Diffusion as experimental and says the demo exists to help develop and refine future models. It sits within the Gemini family, and its benchmark table compares it against Gemini 2.0 Flash-Lite, run through the AI Studio API with default sampling settings.

Inside the model

The model works by refining noise step by step rather than predicting text directly. It emits whole blocks of tokens in one go instead of a single token at a time, which the company says makes replies more coherent and allows error correction during generation. Google DeepMind claims benchmark performance comparable to much larger models while being faster. Reported figures are pass@1 with no majority voting; the SWE-Bench Verified figure is a non-agentic, single-turn edit evaluation with prompts up to 32K.

What it can do

Reported pass@1 scores: LiveCodeBench (v6) 30.9% and SWE-Bench Verified 22.9% on code, GPQA Diamond 40.4% on science, AIME 2025 23.3% on mathematics, BIG-Bench Extra Hard 15.0% on reasoning, and Global MMLU (Lite) 69.1% on multilingual tasks. Against Gemini 2.0 Flash-Lite it scores higher on LiveCodeBench and AIME 2025, and lower on SWE-Bench Verified, GPQA Diamond, BIG-Bench Extra Hard and Global MMLU (Lite).

How to use it

Gemini Diffusion is offered as an experimental demo. Google DeepMind says the demo is there to help develop and refine future models.

Why it matters

Google DeepMind frames diffusion as a new kind of language model worth exploring alongside autoregressive ones, and says the demo will feed into future models.

Questions people ask about Gemini Diffusion

How is Gemini Diffusion different from a normal language model?

Most large language models write text one token at a time, which can be slow. Gemini Diffusion instead learns to produce output by refining noise step by step, so it can emit blocks of tokens together and fix errors while generating.

How fast is it?

Google DeepMind reports an average sampling speed of 1479 tokens per second, excluding overhead, across the evaluations it lists, with 0.84 seconds of overhead. The company says this is significantly faster than its fastest model so far.

How does it compare with Gemini 2.0 Flash-Lite?

On the reported pass@1 scores it is higher on LiveCodeBench (v6) and AIME 2025, and lower on SWE-Bench Verified, GPQA Diamond, BIG-Bench Extra Hard and Global MMLU (Lite). Google DeepMind describes its external benchmark performance as comparable to much larger models.

Can I use it?

It is available as an experimental demo. Google DeepMind says the demo exists to help develop and refine future models.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Gemini Diffusion yet, so the facts on this page come from the sources listed above: deepmind.google/models/gemini-diffusion/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.