son of a bitch blog.google/innovation-a...
June 10, 2026 at 4:10 PM
son of a bitch blog.google/innovation-a...
Most convincing one so far from @mmastrac.bsky.social ! x.com/mmastrac/sta...
September 17, 2026 at 8:17 PM
Most convincing one so far from @mmastrac.bsky.social ! x.com/mmastrac/sta...
DiffusionGemma is fucking fast wow
July 4, 2026 at 2:28 AM
DiffusionGemma is fucking fast wow
DiffusionGemma can now run at 2000+ tokens/sec! ⚡
We made local DiffusionGemma inference 1.8× faster.
Run it on 18GB RAM via Unsloth Studio.
GitHub: github.com/unslothai/un...
Guide: unsloth.ai/docs/models/...
We made local DiffusionGemma inference 1.8× faster.
Run it on 18GB RAM via Unsloth Studio.
GitHub: github.com/unslothai/un...
Guide: unsloth.ai/docs/models/...
June 12, 2026 at 1:57 PM
DiffusionGemma can now run at 2000+ tokens/sec! ⚡
We made local DiffusionGemma inference 1.8× faster.
Run it on 18GB RAM via Unsloth Studio.
GitHub: github.com/unslothai/un...
Guide: unsloth.ai/docs/models/...
We made local DiffusionGemma inference 1.8× faster.
Run it on 18GB RAM via Unsloth Studio.
GitHub: github.com/unslothai/un...
Guide: unsloth.ai/docs/models/...
Diffusion nerds are at it again — DiffusionGemma 26B-A4B
unlike previous language diffusion models, this one doesn’t suck, and it’s very fast
blog.google/innovation-a...
unlike previous language diffusion models, this one doesn’t suck, and it’s very fast
blog.google/innovation-a...
June 10, 2026 at 4:46 PM
Diffusion nerds are at it again — DiffusionGemma 26B-A4B
unlike previous language diffusion models, this one doesn’t suck, and it’s very fast
blog.google/innovation-a...
unlike previous language diffusion models, this one doesn’t suck, and it’s very fast
blog.google/innovation-a...
Want 4x faster local inference on dedicated GPUs for your interactive apps? DiffusionGemma is an experimental, open 26B MoE model that generates entire blocks of text simultaneously instead of token-by-token.
June 10, 2026 at 4:16 PM
Want 4x faster local inference on dedicated GPUs for your interactive apps? DiffusionGemma is an experimental, open 26B MoE model that generates entire blocks of text simultaneously instead of token-by-token.
New gemma!!! And it's a diffusion model! Deepmind keeps releasing diffusion stuff 🤔 it's not that much worse on benches compared to the same sized autoregressive Gemma 4
DiffusionGemma: 4x faster text generation
An overview of DiffusionGemma, an exceptionally fast text generation model with up to 4x faster speeds.
blog.google
June 10, 2026 at 4:13 PM
New gemma!!! And it's a diffusion model! Deepmind keeps releasing diffusion stuff 🤔 it's not that much worse on benches compared to the same sized autoregressive Gemma 4
Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
September 23, 2026 at 12:06 AM
Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
got around to setting up DiffusionGemma on my 5090 - its output is not amazing, but it is incredibly, unbelievably fast
it is *insane* to see 500 tokens per second at home
it is *insane* to see 500 tokens per second at home
August 10, 2026 at 9:22 PM
got around to setting up DiffusionGemma on my 5090 - its output is not amazing, but it is incredibly, unbelievably fast
it is *insane* to see 500 tokens per second at home
it is *insane* to see 500 tokens per second at home
Amazing! Have you tried something like a DiffusionGemma fine tune for this yet?
September 26, 2026 at 7:21 PM
Amazing! Have you tried something like a DiffusionGemma fine tune for this yet?
you need to be looking at DiffusionGemma if you're a curious person imo
Google DeepMind's DiffusionGemma Technical Report
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
August 4, 2026 at 3:16 PM
you need to be looking at DiffusionGemma if you're a curious person imo
DiffusionGemma is so fast. like i can't state how fast. it's kind of nuts
July 4, 2026 at 5:21 PM
DiffusionGemma is so fast. like i can't state how fast. it's kind of nuts
Instead of generating text token by token, "System 1" decision models (like Jev) act as fast, deterministic decision engines. In a single forward pass, they output calibrated probabilities.
Here's how I turned DiffusionGemma into a multimodal visual classifier in Java! 🧵👇
Here's how I turned DiffusionGemma into a multimodal visual classifier in Java! 🧵👇
September 24, 2026 at 7:52 PM
Instead of generating text token by token, "System 1" decision models (like Jev) act as fast, deterministic decision engines. In a single forward pass, they output calibrated probabilities.
Here's how I turned DiffusionGemma into a multimodal visual classifier in Java! 🧵👇
Here's how I turned DiffusionGemma into a multimodal visual classifier in Java! 🧵👇
ok i love DiffusionGemma. i am rooting for more Diffusion language models. but. qwen3.8 at MTP is about as fast and just...better.
August 16, 2026 at 5:42 PM
ok i love DiffusionGemma. i am rooting for more Diffusion language models. but. qwen3.8 at MTP is about as fast and just...better.
What would happen if we tried the diffusion generation on LLMs?
We get Diffusion Gemma!
4x speed up! ⚡⚡⚡💎
blog.google/innovation-a...
We get Diffusion Gemma!
4x speed up! ⚡⚡⚡💎
blog.google/innovation-a...
DiffusionGemma: 4x faster text generation
An overview of DiffusionGemma, an exceptionally fast text generation model with up to 4x faster speeds.
blog.google
June 11, 2026 at 7:34 AM
What would happen if we tried the diffusion generation on LLMs?
We get Diffusion Gemma!
4x speed up! ⚡⚡⚡💎
blog.google/innovation-a...
We get Diffusion Gemma!
4x speed up! ⚡⚡⚡💎
blog.google/innovation-a...
Google DeepMind's DiffusionGemma Technical Report
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
August 4, 2026 at 2:39 PM
Google DeepMind's DiffusionGemma Technical Report
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
2. Chiang claims that the fact that next token selection is the output method of LLMs as if that was some sort of proof of their non-mindedness. It is irrelevant, actually.
The most cutting-edge model publicly released, DiffusionGemma by Google, doesn't primarily use next-token selection.
The most cutting-edge model publicly released, DiffusionGemma by Google, doesn't primarily use next-token selection.
DiffusionGemma
An experimental open model that explores an exceptionally fast approach to text generation
deepmind.google
July 2, 2026 at 11:55 PM
2. Chiang claims that the fact that next token selection is the output method of LLMs as if that was some sort of proof of their non-mindedness. It is irrelevant, actually.
The most cutting-edge model publicly released, DiffusionGemma by Google, doesn't primarily use next-token selection.
The most cutting-edge model publicly released, DiffusionGemma by Google, doesn't primarily use next-token selection.
lmao I saw the diffusiongemma 26B reimplementation, but if that's real then big oof
September 19, 2026 at 6:33 PM
lmao I saw the diffusiongemma 26B reimplementation, but if that's real then big oof
kind of nervous that having Fable touch my DiffusionGemma project is going to sabotage it
July 1, 2026 at 8:39 PM
kind of nervous that having Fable touch my DiffusionGemma project is going to sabotage it
Can the new DiffusionGemma model help fix broken OCR?
In theory, denoising tokens in parallel could work better for OCR correction since context is seen upfront?
Pointed it at 19th-century newspaper OCR. It corrected better than the autoregressive baseline — at ~8x the speed.
In theory, denoising tokens in parallel could work better for OCR correction since context is seen upfront?
Pointed it at 19th-century newspaper OCR. It corrected better than the autoregressive baseline — at ~8x the speed.
June 11, 2026 at 5:03 AM
Can the new DiffusionGemma model help fix broken OCR?
In theory, denoising tokens in parallel could work better for OCR correction since context is seen upfront?
Pointed it at 19th-century newspaper OCR. It corrected better than the autoregressive baseline — at ~8x the speed.
In theory, denoising tokens in parallel could work better for OCR correction since context is seen upfront?
Pointed it at 19th-century newspaper OCR. It corrected better than the autoregressive baseline — at ~8x the speed.
Google releases DiffusionGemma.✨
The new 26B-A4B diffusion text model runs locally on 18GB RAM.
It supports high-speed text generation, thinking, image, video and 256K context.
Run and train via Unsloth Studio.
GGUF: huggingface.co/unsloth/diff...
Guide: unsloth.ai/docs/models/...
The new 26B-A4B diffusion text model runs locally on 18GB RAM.
It supports high-speed text generation, thinking, image, video and 256K context.
Run and train via Unsloth Studio.
GGUF: huggingface.co/unsloth/diff...
Guide: unsloth.ai/docs/models/...
June 10, 2026 at 4:18 PM
Google releases DiffusionGemma.✨
The new 26B-A4B diffusion text model runs locally on 18GB RAM.
It supports high-speed text generation, thinking, image, video and 256K context.
Run and train via Unsloth Studio.
GGUF: huggingface.co/unsloth/diff...
Guide: unsloth.ai/docs/models/...
The new 26B-A4B diffusion text model runs locally on 18GB RAM.
It supports high-speed text generation, thinking, image, video and 256K context.
Run and train via Unsloth Studio.
GGUF: huggingface.co/unsloth/diff...
Guide: unsloth.ai/docs/models/...
This doesn't look like an official release, just a random jev implementation atop diffusiongemma via cloud run?
September 23, 2026 at 9:20 AM
This doesn't look like an official release, just a random jev implementation atop diffusiongemma via cloud run?
capabilities jagged. glm 5.2 is very good at many things. it wasn't good at trying to get DiffusionGemma working. extremely bad at it.
July 4, 2026 at 2:06 PM
capabilities jagged. glm 5.2 is very good at many things. it wasn't good at trying to get DiffusionGemma working. extremely bad at it.
DiffusionGemma generates 256 tokens per step and can rewrite any of them mid-generation. #GenerativeAI
August 20, 2026 at 3:30 PM
DiffusionGemma generates 256 tokens per step and can rewrite any of them mid-generation. #GenerativeAI