DiffusionGemma 26B A4B
Google DiffusionGemma 26B A4B: discrete diffusion model denoising 256-token canvases in parallel. Gemma 4 MoE with bidirectional attention and
Model Overview
Capabilities, design details, and architectural traits
DiffusionGemma 26B A4B - Discrete Diffusion Text Generation on a Gemma 4 MoE Backbone
DiffusionGemma replaces token-by-token autoregression with discrete text diffusion. Instead of predicting one token at a time, it starts with a canvas of random placeholder tokens and iteratively denoises them in parallel, shifting the inference bottleneck from memory bandwidth to compute.
| Trait | Detail |
|---|---|
| Generation method | Block-autoregressive multi-canvas sampling: a 256-token canvas is denoised in parallel via a diffusion sampler, committed to the KV cache, then the next canvas begins |
| Architecture | Encoder-decoder design: an autoregressive encoder processes and caches the prompt, while the decoder applies bidirectional attention over the generation canvas via cross-attention |
| Self-correction | Bidirectional context lets every canvas position attend to all others simultaneously, enabling real-time error correction and re-noising when confidence drops |
| Speed profile | Optimized for small-batch inference; parallel denoising of 256 tokens yields 15-20 tokens per forward pass |
| MoE design | 8 active experts out of 128 total plus 1 shared, for low-memory local execution |
Uniform State Diffusion and Block Autoregression
The model uses Uniform State Diffusion: over multiple denoising passes, highly confident tokens help resolve adjacent positions, causing the full sequence to converge. For outputs longer than 256 tokens, the block-autoregressive mechanism commits each finished canvas to the KV cache before initializing a fresh one conditioned on prior history, combining parallel block speed with sequential stability.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Anthropic | Anthropic | |
| Release Date | June 10, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 256K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | Free Best Input Pricing | $4 | $2 |
| Output Pricing | Free Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimagevideo | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 9.5 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 19.7 | - | - |
| Agentic Index | 2.2 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Google
Other models by Google
Gemini 4 Argon
Gemini 3.8 Flash
Gemini 3.7 Flash
Gemini 3.6 Flash
Top AI Models
Leading alternatives by intelligence score
