DiffusionGemma suits teams building speed-sensitive local or on-device apps, especially document parsing and OCR, where its bidirectional decoder attention beats standard causal-attention models. The tradeoff: it trails its own autoregressive backbone, Gemma 4, by 5 to 19 points on reasoning benchmarks, so accuracy-critical coding or math work should stick with that sibling instead.
DiffusionGemma is Google DeepMind's open-source discrete text diffusion language model, released June 10, 2026. Instead of predicting one token at a time, it denoises 256-token blocks of text in parallel, running at over 1,000 tokens per second on modern GPUs, with weights free to self-host commercially.
Provider: Google · Family: DiffusionGemma
Context window: 256,000 tokens
Input modalities: text, image, video, tool-calls · Output: text, tool-calls
About DiffusionGemma 26B-A4B
DiffusionGemma 26B-A4B is Google DeepMind's first discrete text diffusion language model, released publicly on June 10, 2026 as an open-weight model. It runs on the same Mixture-of-Experts backbone as Gemma 4 26B-A4B (128 fine-grained experts, top-8 routing, 25.2 billion total parameters with 3.8 billion active per token), but replaces one-token-at-a-time decoding with a parallel denoising process that generates a full block of tokens per forward pass over up to 48 refinement steps. The design targets speed-critical, low-concurrency workloads: inline editing, rapid prototyping, and interactive assistants where raw throughput matters more than chasing the top benchmark score. On quality benchmarks, DiffusionGemma scores 77.6% on MMLU Pro, 73.2% on GPQA Diamond, 69.1% on AIME 2026 (no tools), and 69.1% on LiveCodeBench v6, with a Codeforces ELO of 1,429. Its autoregressive sibling Gemma 4 26B-A4B beats it on every one of those: 82.6% MMLU Pro, 82.3% GPQA Diamond, 88.3% AIME 2026, a gap of 5 to 19 points depending on the test. The one benchmark DiffusionGemma wins is OmniDocBench 1.5 document parsing, where bidirectional decoder attention gives it a structural edge over causal models at OCR and layout-aware extraction. HumanEval pass@1 with no tools lands at just 11.0%, a known weak spot tied to the block-generation approach struggling with strict left-to-right output formatting. DiffusionGemma reaches 1,008 tokens per second running FP8 precision, and 1,288 tokens per second on an H200, roughly 4x an equivalently-sized autoregressive model on the same class of hardware at low concurrency. vLLM added native support on day one, the first discrete diffusion LLM to get that treatment, with same-day support in Hugging Face Transformers, MLX, and SGLang. That speed edge is specific to low-concurrency use: in high-QPS cloud serving, autoregressive models can batch requests to saturate compute, which narrows or erases the throughput gap.
Pricing
Open weights, so there's no license fee or per-token charge from Google, only the compute you run it on. Self-hosted cost runs about $0.50/hr on a consumer GPU (18 GB VRAM, NVFP4, 700+ tok/s) up to $2.50/hr on an H100 (28 GB VRAM, FP8, 1,008 tok/s). NVIDIA NIM offers free hosted inference in preview. On an H100, that works out to roughly $0.70 per million tokens processed.
Key Features
- Parallel Block Generation: Generates an entire block of tokens through iterative denoising instead of one token at a time, the core mechanism behind its throughput advantage over standard autoregressive decoding.
- 256,000-Token Context Window: Supports a large context window using a bidirectional encoder-decoder design, enabling full-document analysis and multi-document reasoning in a single request.
- Native Multimodal Input: Accepts text, images, and video in a single prompt, with confirmed support for OCR, PDF parsing, chart comprehension, screen parsing, and video analysis.
- Open Weights on Hugging Face: Weights are available on Hugging Face, Kaggle, and Google Cloud Vertex AI Model Garden under an Apache 2.0 license, with no royalties or attribution required for commercial use.
- Runs on a Single Consumer GPU: The NVFP4-quantized build fits within 18 GB of VRAM, putting it within reach of a single high-end consumer GPU rather than requiring a datacenter card.
- Configurable Thinking Mode: Offers a reasoning mode that can be toggled per request, adding step-by-step chain-of-thought before the final answer on math, code, and complex multi-step tasks.
Pros
- The fastest open-source inference speed available in its parameter class: roughly 4x an equivalently-sized autoregressive model at low concurrency.
- Fully open-source: self-host, fine-tune, or deploy commercially with no royalties or attribution requirements.
- Wins on document-parsing and OCR benchmarks thanks to the bidirectional attention baked into its decoder, unlike standard causal-attention models.
- Small enough to run on a single high-end consumer GPU thanks to NVFP4 quantization, unusual for a 25B+ parameter model.
Cons
- Trails its autoregressive sibling by double-digit points on GPQA Diamond, AIME, and MMLU Pro reasoning benchmarks.
- HumanEval pass@1 without tools is a known weak point, reflecting the block-generation approach's trouble with strict left-to-right output formats.
- No audio input or output; full-precision BF16 needs 50+ GB of VRAM across multiple GPUs.
- Speed advantage shrinks at high concurrency: autoregressive models close the gap through request batching in production serving.
Benchmarks
- mmlu pro: 77.6
- aime 2025: 69.1
- humaneval: 11
- live bench: 69.1
- gpqa diamond: 73.2
- artificial analysis speed tokens per sec: 1008
Frequently Asked Questions
How much does DiffusionGemma 26B-A4B cost in 2026?
DiffusionGemma costs nothing to license: it's fully open-weight, so Google charges no per-token fee and you only pay for the compute you run it on. Self-hosting the NVFP4 build costs roughly $0.50 an hour on a consumer GPU, while the faster FP8 tier costs about $2.50 an hour on a datacenter-class card. NVIDIA also runs a free hosted preview through NVIDIA NIM if you'd rather not manage a GPU yourself.
Is DiffusionGemma 26B-A4B free to use?
Yes. It ships fully open-weight, so self-hosting has no usage caps, seat limits, or per-token fees. NVIDIA's hosted NIM endpoint is also free to try, though capacity is limited during its preview.
What are the best alternatives to DiffusionGemma 26B-A4B?
DeepSeek V4 offers higher benchmark accuracy if you're running at scale and don't need the throughput advantage. Mistral Medium 3 is a good fit if you'd rather use a low-cost hosted API instead of self-hosting. Both trade DiffusionGemma's inference speed and open licensing for either stronger reasoning scores or zero infrastructure management.
How does DiffusionGemma 26B-A4B compare to Gemma 4 26B-A4B?
Gemma 4, the autoregressive backbone DiffusionGemma is built from, wins on every general reasoning benchmark: MMLU Pro (82.6% vs 77.6%), GPQA Diamond (82.3%, roughly a nine-point gap), and AIME 2026 (88.3% vs 69.1%). DiffusionGemma's advantage is speed, not accuracy: it generates whole blocks of tokens in parallel instead of one at a time, roughly 4x faster at low concurrency on the same GPU. The one benchmark where DiffusionGemma pulls ahead is document parsing, where its bidirectional attention outperforms Gemma 4 on OCR and layout-heavy extraction.
How do you get started with DiffusionGemma 26B-A4B?
Download the weights from Hugging Face and run them with vLLM, which added native support for DiffusionGemma at launch. Pick your precision based on available VRAM: the NVFP4 build is the easiest starting point for a single consumer GPU, while FP8 needs a datacenter-class card. If you'd rather not manage infrastructure, NVIDIA NIM's free hosted preview lets you test the model with an API key and no local setup.
Top Alternatives
- DeepSeek V4: Pick DeepSeek V4 for higher benchmark accuracy at scale; pick DiffusionGemma for faster, self-hosted open-weight inference.
- Mistral Medium 3: Pick Mistral Medium 3 for a low-cost hosted API; pick DiffusionGemma for self-hosted inference at roughly 4x the throughput.