DiffusionGemma
Hacker ApprovedRanked 2 of 3 diffusion language models by recommenders
Diffusion-based open model people recommend for very fast token output, especially on a single GPU. Common caveat: output quality often trails regular autoregressive Gemma models.
Diffusion-based open model people recommend for very fast token output, especially on a single GPU. Common caveat: output quality often trails regular autoregressive Gemma models.
22 people gave an opinion
- recommend it
- 15
- criticize it
- 7
- approval share
- 68%
Recommended in 5 of 13 months
- November 2025: not recommended
- December 2025: not recommended
- January 2026: not recommended
- February 2026: not recommended
- March 2026: not recommended
- April 2026: not recommended
- May 2026: not recommended
- June 2026: recommended
- July 2026: recommended
- August 2026: recommended
- September 2026: recommended
- October 2026: recommended
- November 2026: not recommended (month in progress)
Why it carries the badge
- 15 people recommend itNeeds 10 or more, met.
- 2.1 recommenders for every criticNeeds 2 to 1 or better, met.
- Recommended in 5 of 13 monthsNeeds 4 or more, met.
The same three rules apply to every product, and the badge can’t be bought. How the badge works
What people say about it
The reasons people give, with the number of comments that give each one. Open a reason to read what they said.
Why people recommend it
Each mark is one comment
Very fast token generation
“Instead, diffusion models work much faster for individual prompts, and not by a small margin either.”
Comment from 16 Jun 2026 in “Running local models is good now” “diffusion is huge for local MoE inference because of the improved token generation efficiency”
Comment from 11 Jun 2026 in “DiffusionGemma: 4x Faster Text Generation” “sadly isn't as smart as Gemma itself, but holy hell is it FAST”
Comment from 27 Jul 2026 in “Kimi-K3 Releases on HuggingFace 6/27” - Comment from 20 Aug 2026 in “DiffusionGemma Technical Report”
- Comment from 13 Jun 2026 in “Open Source AI Must Win”
- Comment from 11 Aug 2026 in “Antirez/h3.c: MiniMax H3 inference engine for Mac computers”
- Comment from 16 Jun 2026 in “Running local models is good now”
- Comment from 31 Aug 2026 in “How to build a diffusion language model”
- Comment from 3 Jul 2026 in “Jamesob's guide to running SOTA LLMs locally”
Fast on a single GPU
“single-user scenarios, which is where these diffusion models really make a large difference in performance.”
Comment from 16 Jun 2026 in “Running local models is good now” “>900TPS output reliably on a single 5090 with the nvfp4 model.”
Comment from 20 Aug 2026 in “DiffusionGemma Technical Report” “it really makes good use of single GPUs for single responses”
Comment from 20 Aug 2026 in “DiffusionGemma Technical Report” - Comment from 11 Aug 2026 in “Antirez/h3.c: MiniMax H3 inference engine for Mac computers”
- Comment from 31 Aug 2026 in “How to build a diffusion language model”
Good responses for its size
“seems to perform better (in responses) than benchmarks seemed to indicate and reasonably fast for its model size”
Comment from 9 Jul 2026 in “Getting GLM 5.2 running on my slow computer” “I like the model a lot and it's fairly good at reasoning.”
Comment from 20 Aug 2026 in “DiffusionGemma Technical Report”
What people warn about
Each mark is one comment
Output quality trails autoregressive models
“I'd still eat the time cost of the normal Gemma for 10% extra performance”
Comment from 10 Jun 2026 in “DiffusionGemma: 4x Faster Text Generation” “performed a lot worse than its equivalent autoregressive counterpart”
Comment from 19 Oct 2026 in “Mercury 3: Diffusion at scale” “sadly isn't as smart as Gemma itself, but holy hell is it FAST”
Comment from 27 Jul 2026 in “Kimi-K3 Releases on HuggingFace 6/27” - Comment from 12 Jun 2026 in “DiffusionGemma: Discrete diffusion in a large language model”
Large drops on hard benchmarks
“Look at the dgemma benchmarks - pretty steep dropoffs and the more difficult the benchmark the worse the dropoff.”
Comment from 16 Jun 2026 in “Running local models is good now” “Large dropoffs, with the hardest benchmarks seeing the largest drops.”
Comment from 10 Jun 2026 in “DiffusionGemma: 4x Faster Text Generation”
Other diffusion language models people recommend
The two most recommended other diffusion language models, with the same counts for each, so you can compare like for like. The tick on each bar marks the 2 to 1 line the badge needs.
| Product | Recommend | Criticize | Approval | Months | Owned | Badge |
|---|---|---|---|---|---|---|
| Inception Mercury 2 | 64 | 28 | 70% | 6 | — | Approved |
| DiffusionGemmaThis page | 15 | 7 | 68% | 5 | — | Approved |
| Inception Labs | 5 | 1 | 83% | 4 | — | — |
What people said
31 of the 32 comments that mention it, newest first. Each links to the original comment.
“I'm happy to see efforts like Mercury 3 and Diffusion Gemma.”
“gemma diffusion can be repurposed as a decision model with output tokens and diffusion steps set to 1”
“it barely gets past 5 words before it completely gets it wrong”
“We're sleeping on a amazing model right in front of our eyes.”
“performed a lot worse than its equivalent autoregressive counterpart”
“Diffusion Gemma sucks to train if you are not already mostly in distribution”
“Things that were Qwen or Gemma Diffusion did better at those kinds of tasks but were generally pretty inconsistent”
“DiffusionGemma has a lot of knowledge that may be challenging to train into the smaller models”
“the real weakness of diffusion models: you can have two positions where two tokens need to be coordinated”
“it is very fast on GPUs in output token/sec”
“>900TPS output reliably on a single 5090 with the nvfp4 model.”
“diffusion is a bad fit for metal”
“I re-implemented this one for macOS over the last couple of months”
“I like the model a lot and it's fairly good at reasoning.”
“it really makes good use of single GPUs for single responses”
“holy moly is it FAST on one GPU, seriously impressive”
“sadly isn't as smart as Gemma itself, but holy hell is it FAST”
“seems to perform better (in responses) than benchmarks seemed to indicate and reasonably fast for its model size”
“DiffusionGemma is around 4x regular gemma.”
“single-user scenarios, which is where these diffusion models really make a large difference in performance.”
“DiffusionGemma seem to both be very high token rate”
“Look at the dgemma benchmarks - pretty steep dropoffs and the more difficult the benchmark the worse the dropoff.”
“Instead, diffusion models work much faster for individual prompts, and not by a small margin either.”
“the root issue is that DiffusionGemma emits incorrect JSON”
“Algorithmic advances like DiffusionGemma ~4x token gen speeds”
“Impressive speed up at the cost of quality”
“diffusion is huge for local MoE inference because of the improved token generation efficiency”
“it requires a fork of llama.cpp to run at the moment”
“I'm not getting anywhere near the speeds advertised on my 3090 Ti, alas, but it's fun watching it "fill out" its answers.”
“I'd still eat the time cost of the normal Gemma for 10% extra performance”
“Large dropoffs, with the hardest benchmarks seeing the largest drops.”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.