{"slug": "diffusiongemma", "name": "DiffusionGemma", "category": "ai-models", "category_name": "AI models", "description": "", "recommenders": 5, "commenters": 8, "warners": 0, "score": 4.699, "rising": 2.35, "first": "2026-06-10", "latest": "2026-09-27", "why": "I like the model a lot and it's fairly good at reasoning.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 6, "neutral": 3, "compare": 1, "warn": 0}, "spark": [0, 0, 0, 0, 0, 0, 0, 2, 0, 3, 1, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "I like the model a lot and it's fairly good at reasoning.", "kind": "", "icon": false, "family": "gemma", "domain": "deepmind.google", "github_repo": null, "official_url": "", "pricing": "", "aliases": ["DiffusionGemma"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 795753936, "item_id": 49734621, "url": "https://news.ycombinator.com/item?id=49734621", "by": "mmastrac", "date": "2026-09-16", "version": null, "stance": "recommend", "quote": "It (DiffusionGemma with the Jev mode) can also solve an ASCII maze.", "excerpt": "Any diffusion model is potentially a Jev in disguise: https://github.com/vllm-project/vllm/pull/57250 Runs ~0.2s per decision on my DGX Spark. 10/10 programming language detection 9/10 human language detection 10/12 unit magnitude comparison All incorrect answers are marked with low-P. It (DiffusionGemma with the Jev mode) can also solve an ASCII maze.", "context": "runs it on a DGX Spark", "use_case": "fast on-device decisions", "switched_from": "", "story": "Reverse-engineered Jev-like model"}, {"id": 790030400, "item_id": 49376900, "url": "https://news.ycombinator.com/item?id=49376900", "by": "mmastrac", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "I like the model a lot and it's fairly good at reasoning.", "excerpt": "I re-implemented this one for macOS over the last couple of months: https://github.com/mmastrac/diffgemma I like the model a lot and it's fairly good at reasoning. You can also really bend it to your needs. It's designed for machines with more compute than memory bandwidth but IMO does really well on metal. I've got it up to ~15tok/s on M3-class machines, but I wager there's a bunch of perf on M5 that I just don't ha", "context": "", "use_case": "local inference on macOS", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 790000448, "item_id": 49375028, "url": "https://news.ycombinator.com/item?id=49375028", "by": "embedding-shape", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "Probably the fastest model I've been able to run so far", "excerpt": "DiffusionGemma goes one step further even, and does this denoising over multiple \"canvases\" which lets it do reasoning and separate out a \"final reply\" canvas, looks something like this: https://gist.github.com/embedding-shapes/f4cb46bad704b6d016888c28e0572f2e Diffusion text models for me is the more interesting type of LLMs for local usage, as it really makes good use of single GPUs for single responses, rather than", "context": "", "use_case": "Local text generation", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 789996848, "item_id": 49374803, "url": "https://news.ycombinator.com/item?id=49374803", "by": "kamranjon", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "What makes me hopeful about this release is that possibly this same conversion can be applied", "excerpt": "Just wanted to share this, I found it was a really nice resource to understand how diffusion Gemma worked: https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-diffusiongemma The really interesting thing to me was that they didn\u2019t need to train this model from scratch they just used their existing MOE checkpoint: \u201cTo convert a decoder-only model (Gemma 4 26B A4B) into a denoiser, we can make use of somethin", "context": "", "use_case": "", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 776970066, "item_id": 48560629, "url": "https://news.ycombinator.com/item?id=48560629", "by": "jauntywundrkind", "date": "2026-06-16", "version": null, "stance": "recommend", "quote": "DiffusionGemma seem to both be very high token rate", "excerpt": "i'd love to get to a point where big models can launch subagents that are fast and local. there's a lot of focus on token rate, but just as much, it feels like there's other latencies that are there with these big batched processing runs, and i think local might have some real wins. Gemma 4 seems already on the right track. lfm2.5-8b-a1b (https://www.liquid.ai/blog/lfm2-5-8b-a1b) and DiffusionGemma seem to both be ve", "context": "", "use_case": "fast local subagents", "switched_from": "", "story": "Running local models is good now"}]}], "caveats": [], "who": [{"by": "mmastrac", "context": "runs it on a DGX Spark", "url": "https://news.ycombinator.com/item?id=49734621", "date": "2026-09-16"}], "tags": [], "use_cases": ["evals", "fast on-device decisions", "local inference on macos", "local text generation", "experimenting with reasoning control", "fast local subagents"], "compared": [{"slug": "qwen", "name": "Qwen", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "gemma", "name": "Gemma", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "jev", "name": "Jev", "co": 3, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=49374287", "title": "DiffusionGemma Technical Report", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=49731282", "title": "Reverse-engineered Jev-like model", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=49861274", "title": ">1B tokens/minute/GPU by combining query planner and inference engine", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49752041", "title": "OpenJev", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48979475", "title": "Controlling Reasoning Effort in LLMs", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48555993", "title": "Running local models is good now", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {}, "2026-04": {}, "2026-05": {}, "2026-06": {"neutral": 1, "recommend": 2}, "2026-07": {"neutral": 1}, "2026-08": {"recommend": 3}, "2026-09": {"recommend": 1, "compare": 1, "neutral": 1}, "2026-10": {}}, "versions": [], "version_unnamed": 8, "family_ref": {"slug": "gemma", "name": "Gemma"}, "runs_on": null, "lines": [], "quotes": [{"id": 797789857, "item_id": 49861866, "url": "https://news.ycombinator.com/item?id=49861866", "by": "charles_irl", "date": "2026-09-27", "version": null, "stance": "neutral", "quote": "We have some more results on DiffusionGemma", "excerpt": "Thanks! We chose the Qwen model for the initial and core results because it's a simple architecture, which made the roofline cost modeling very easy and trustworthy. I'd already done related work with that model before for another project on speculative decoding, so we had a reference impl as well. For brevity, this post focuses on a subset of the results. In the companion post on the Full Stack Data Lab blog[1], we ", "context": "", "use_case": "", "switched_from": "", "story": ">1B tokens/minute/GPU by combining query planner and inference engine"}, {"id": 796093970, "item_id": 49755873, "url": "https://news.ycombinator.com/item?id=49755873", "by": "mmastrac", "date": "2026-09-18", "version": null, "stance": "compare", "quote": "DG wins some, Jev wins some, both show low confidence when wrong", "excerpt": "If you want to try a _legit_ Jev implementation that matches (at least in my evals), the vLLM patch to turn DiffusionGemma into Jev is available. On my DGX Spark I get very similar latency numbers, and it matches my evals + or - a few points on each test (DG wins some, Jev wins some, both show low confidence when wrong). I ran the same evals against a Qwen36 and it clearly lost to both of them, so you are leaving bot", "context": "runs on a DGX Spark", "use_case": "evals", "switched_from": "", "story": "OpenJev"}, {"id": 795901665, "item_id": 49743854, "url": "https://news.ycombinator.com/item?id=49743854", "by": "budro", "date": "2026-09-17", "version": null, "stance": "compare", "quote": "DiffusionGemma has a vision tower and significantly more spatial data in its training set", "excerpt": "Your maze demo gave me the idea to try out the \"speculative fan-out\" pattern [1]. It seemed interesting to try solving mazes in one-shot. Unfortunately it seems like Jev can't reliably solve basic mazes even with a step count of 1! [2] I was very surprised. Could you point me in the direction of your maze solving code so I can see if it's a skill issue? The only other explanation I can come up with is that Jev was no", "context": "", "use_case": "", "switched_from": "", "story": "Reverse-engineered Jev-like model"}, {"id": 795753936, "item_id": 49734621, "url": "https://news.ycombinator.com/item?id=49734621", "by": "mmastrac", "date": "2026-09-16", "version": null, "stance": "recommend", "quote": "It (DiffusionGemma with the Jev mode) can also solve an ASCII maze.", "excerpt": "Any diffusion model is potentially a Jev in disguise: https://github.com/vllm-project/vllm/pull/57250 Runs ~0.2s per decision on my DGX Spark. 10/10 programming language detection 9/10 human language detection 10/12 unit magnitude comparison All incorrect answers are marked with low-P. It (DiffusionGemma with the Jev mode) can also solve an ASCII maze.", "context": "runs it on a DGX Spark", "use_case": "fast on-device decisions", "switched_from": "", "story": "Reverse-engineered Jev-like model"}, {"id": 790030400, "item_id": 49376900, "url": "https://news.ycombinator.com/item?id=49376900", "by": "mmastrac", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "I like the model a lot and it's fairly good at reasoning.", "excerpt": "I re-implemented this one for macOS over the last couple of months: https://github.com/mmastrac/diffgemma I like the model a lot and it's fairly good at reasoning. You can also really bend it to your needs. It's designed for machines with more compute than memory bandwidth but IMO does really well on metal. I've got it up to ~15tok/s on M3-class machines, but I wager there's a bunch of perf on M5 that I just don't ha", "context": "", "use_case": "local inference on macOS", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 790000448, "item_id": 49375028, "url": "https://news.ycombinator.com/item?id=49375028", "by": "embedding-shape", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "Probably the fastest model I've been able to run so far", "excerpt": "DiffusionGemma goes one step further even, and does this denoising over multiple \"canvases\" which lets it do reasoning and separate out a \"final reply\" canvas, looks something like this: https://gist.github.com/embedding-shapes/f4cb46bad704b6d016888c28e0572f2e Diffusion text models for me is the more interesting type of LLMs for local usage, as it really makes good use of single GPUs for single responses, rather than", "context": "", "use_case": "Local text generation", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 789996848, "item_id": 49374803, "url": "https://news.ycombinator.com/item?id=49374803", "by": "kamranjon", "date": "2026-08-20", "version": null, "stance": "recommend", "quote": "What makes me hopeful about this release is that possibly this same conversion can be applied", "excerpt": "Just wanted to share this, I found it was a really nice resource to understand how diffusion Gemma worked: https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-diffusiongemma The really interesting thing to me was that they didn\u2019t need to train this model from scratch they just used their existing MOE checkpoint: \u201cTo convert a decoder-only model (Gemma 4 26B A4B) into a denoiser, we can make use of somethin", "context": "", "use_case": "", "switched_from": "", "story": "DiffusionGemma Technical Report"}, {"id": 783760400, "item_id": 48985025, "url": "https://news.ycombinator.com/item?id=48985025", "by": "embedding-shape", "date": "2026-07-20", "version": null, "stance": "neutral", "quote": "gonna give it a try with DiffusionGemma", "excerpt": "> My favorite trick for controlling the reasoning level is the hack First described here I think, specifically for Qwen models: https://arxiv.org/abs/2501.19393, although it seems unclear how much it actually improves the final response unless you create a new fine-tuned model with this in mind, and also if it's possible to replicate for other models. Still, got inspired and gonna give it a try with DiffusionGemma al", "context": "", "use_case": "experimenting with reasoning control", "switched_from": "", "story": "Controlling Reasoning Effort in LLMs"}, {"id": 776970066, "item_id": 48560629, "url": "https://news.ycombinator.com/item?id=48560629", "by": "jauntywundrkind", "date": "2026-06-16", "version": null, "stance": "recommend", "quote": "DiffusionGemma seem to both be very high token rate", "excerpt": "i'd love to get to a point where big models can launch subagents that are fast and local. there's a lot of focus on token rate, but just as much, it feels like there's other latencies that are there with these big batched processing runs, and i think local might have some real wins. Gemma 4 seems already on the right track. lfm2.5-8b-a1b (https://www.liquid.ai/blog/lfm2-5-8b-a1b) and DiffusionGemma seem to both be ve", "context": "", "use_case": "fast local subagents", "switched_from": "", "story": "Running local models is good now"}, {"id": 776211538, "item_id": 48513221, "url": "https://news.ycombinator.com/item?id=48513221", "by": "verdverm", "date": "2026-06-12", "version": null, "stance": "recommend", "quote": "Algorithmic advances like DiffusionGemma ~4x token gen speeds", "excerpt": "DGX Spark runs this sized model (I personally like qwen36moe better than gemma4moe) at speeds fast enough for interactive coding sessions. Algorithmic advances like DiffusionGemma ~4x token gen speeds (https://deepmind.google/models/gemma/diffusiongemma/)", "context": "", "use_case": "", "switched_from": "", "story": "Open Source AI Must Win"}, {"id": 775687696, "item_id": 48480481, "url": "https://news.ycombinator.com/item?id=48480481", "by": "simonw", "date": "2026-06-10", "version": null, "stance": "neutral", "quote": "NVIDIA are hosting a free endpoint for this one", "excerpt": "NVIDIA are hosting a free endpoint for this one, details at https://build.nvidia.com/google/diffusiongemma-26b-a4b-it - you have to create an account and (I think) verify a phone number too. (I got it to draw a pelican: https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fe5e234a6dc6eef61e209ce1629620042 )", "context": "", "use_case": "Text generation model", "switched_from": "", "story": "DiffusionGemma: 4x Faster Text Generation"}]}