diffusiongemma-26B-A4B-it

Benchmarks 15 rows

Summary generated, cited

This Mixture-of-Experts architecture uses discrete diffusion to generate text from text, image, and video inputs.^[1][2]

1. is mixture of experts
2. “generates tokens using discrete diffusion”

Generated 2026-10-07 15:02 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

What it is good at 8 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

• Reason through hard problemsauthor-reportedAuthor reports MMLU Pro 77.6%Author reports AIME 2026 no tools 69.1%Author reports GPQA Diamond 73.2%+4 more
• Understand images and videoauthor-reportedAuthor reports MMMU Pro 54.3%Author reports MATH-Vision 70.5%decision-model estimate p=0.87+3 more
• Call tools and APIsauthor-reportedAuthor reports Tau2 (average over 3) 56.2%chat template: tool_calls
• Read long documentsauthor-reportedAuthor reports MRCR v2 8 needle 128k (average) 32.0%config: 262,144 context
• Write code with agentsauthor-reportedAuthor reports LiveCodeBench v6 69.1%Author reports Codeforces ELO 1429
• Work across languagesauthor-reportedAuthor reports MMMLU 81.5%
• Extract text from documentsauthor-reportedAuthor reports OmniDocBench 1.5 (average edit distance, lower is better) 0.319
• Chat and follow instructionsstructuraldecision-model estimate p=0.96name: ittag: conversational+1 more

Facts

• diffusiongemma-26B-A4B-it has 25.8B parameters (4B active per token). (safetensors metadata, observed 2026-10-24)
• diffusiongemma-26B-A4B-it is a mixture-of-experts model with 128 experts. (config.json, observed 2026-10-24)
• Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
• License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
• Created 2026-06-09, last modified 2026-07-15. (HF API, observed 2026-10-24)

Servability

vLLM: supported · DiffusionGemmaForBlockDiffusion · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit ESTIMATE

Weights-only memory for diffusiongemma-26B-A4B-it: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

Lineage

quantized: 37finetune: 37adapter: 8merge: 1

Top variants (40 shown, ranked by downloads)

Repository changelog 4 commits

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Similar models

• gemma-4-26B-A4B-it · 25.8B
  same size band (L), also MoE, same task (image-text-to-text), more providers, permissive license
• gemma-4-26B-A4B · 25.8B
  same size band (L), also MoE, same task (image-text-to-text), more providers, permissive license
• Qwen3-VL-30B-A3B-Instruct · 31.1B
  same size band (L), also MoE, same task (image-text-to-text), more providers, permissive license

Sources

• Hugging Face model repo
• HF Hub API (full)
• hub-stats parquet

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:02 UTC · Last verified 2026-10-25