AI Launch Profile
DiffusionGemma
Google DeepMind released DiffusionGemma, an experimental Apache-2.0 open-weights 25.2B mixture-of-experts model that generates text by iteratively denoising 256-token canvases in parallel.
At a glance
Launch Snapshot
- Company
- Google DeepMind
- Launch date
- June 10, 2026
- Launch type
- Not classified
- Category
- AI Models, AI Open-Weight Models
- Audience
- AI Engineers, Developers
- Pricing
- The weights are Apache-2.0 licensed with no model license fee. Self-hosted compute, storage, networking, monitoring and engineering costs apply; hosted deployment costs vary by provider.
- Free plan
- Yes
- API
- Yes
- Open weights/source
- Yes
Launch Context
Use these links to move from this record into the broader Launch Intelligence database.
Verification & Sources
- Evidence state
- Recheck due
- Source links
- 5
- Freshness
- Needs recheck: checked July 28, 2026
- Last updated
- July 28, 2026
What this evidence state means
- Definition
- The claim was previously checked, but its review window expired or a material change may have invalidated it.
- Required provenance
- The prior evidence and check date are retained, together with the expiry or change signal that triggered recheck.
- Owner
- Kingy freshness queue owner and assigned editorial reviewer
- Freshness rule
- This is already outside its freshness rule. It must not be presented as current until reviewed against current evidence.
- Disputes and corrections
- Use “Suggest a correction” on the record. Kingy editorial reviews the cited evidence, records material corrections, and changes or removes the state when it is not supported.
Key source checks
Suggest a correction
Kingy Launch Score
8.3 / 10 · Solid
One earned credibility score, computed from cited evidence — not a placeholder. How the Kingy Launch Score works
Why this score
- Source & verification 8.5/10 — Official model page, developer guide, Hugging Face model card, and model-card index — four dated first-party sources. deepmind.google
- Product evidence 9.0/10 — Open weights on Hugging Face plus a developer guide — downloadable and directly testable. huggingface.co
- Significance & novelty 8.0/10 — A 26B mixture-of-experts open model generating token blocks through diffusion rather than sequential decoding — technically novel, explicitly experimental. developers.googleblog.com
- Traction signals not scored — insufficient sourced evidence
- Offer clarity 7.0/10 — Published as an experimental open model under the applicable Gemma terms; deployment costs depend on your infrastructure. deepmind.google
Evidence checked 2026-07-10
Kingy AI Take
DiffusionGemma is a credible architecture experiment for small-batch GPU generation, but Google’s own benchmark table shows substantial quality trade-offs against Gemma 4 on many tasks. Kingy did not run the model, reproduce the advertised throughput or validate safety claims, so adoption should depend on a paired hardware, quality and cost trial.
Who it is for
Model-serving researchers, developers and AI engineering teams evaluating low-latency text generation on capable GPUs.
What feels promising
Parallel canvas denoising can use GPU compute differently from token-by-token decoding, with vLLM support lowering the serving barrier.
What feels unproven
Kingy did not reproduce speed, benchmark, memory, long-context, multimodal or safety results on the intended deployment stack.
Source list
Sources
Related Kingy Links
Editorial submissions and sponsor-fit reviews are separate. Payment does not influence Kingy scores, verdicts, rankings, evidence labels, or publication decisions.