Up to date
google Gemma Terms of Use out 23 months ago
3.03B parameters, images and text in, text out.
Verified 2026-10-25 Rank #582 in LLMs · score 74.3 method v1 Sources 3 Downloads 30d 6.7k · #528 of 619 tracked
Picked by rule from 3 tracked pieces: the lab's own post, the most shared, a field report, the latest.
Google launched PaliGemma 2 mix checkpoints in February 2025, offering a vision-language model tuned for multiple tasks like captioning and OCR.
The 1 pick above, plus these 2. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 2 quantized builds and 4 Spaces.
Weights only, an estimate: leave room for the KV cache. Method
About 1.8 GB at INT4, 3.6 GB at FP8, 7.3 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2024-11-21), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.
235 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported [COCO-35L][coco-35l] (avg34) | 115.8 | 448-3B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (avg34) | 117.2 | 448-10B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (avg34) | 117.2 | 448-28B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 138.4 | 224-3B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 140.8 | 224-10B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 142.4 | 224-28B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 140.4 | 448-3B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 142.4 | 448-10B | model author | 2026-10-07 |
| author-reported [COCO-35L][coco-35l] (en) | 142.3 | 448-28B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 141.3 | 224-3B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 143.7 | 224-10B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 144 | 224-28B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 143.4 | 448-3B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 145 | 448-10B | model author | 2026-10-07 |
| author-reported [COCOcap][coco-cap] | 145.2 | 448-28B | model author | 2026-10-07 |
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.
Listed in LLMs
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | PaliGemmaForConditionalGeneration |
| Model type | paligemma |
| Dtype | bfloat16 |
| Params by dtype | BF16: 3,032,242,416 |
| Size band | M |
| MoE | no |
| Gated | manual |
| Papers | arXiv:2407.07726, arXiv:2408.00118, arXiv:2310.09199, arXiv:2303.15343, arXiv:1706.03762, arXiv:2010.11929, arXiv:2412.03555, arXiv:2209.06794, arXiv:2209.04372, arXiv:2103.01913, arXiv:1908.04913, arXiv:1906.02467, arXiv:2203.10244, arXiv:2205.12522, arXiv:2104.12756, arXiv:1608.00272, arXiv:1511.02283, arXiv:1905.13648, arXiv:2110.11624, arXiv:2108.03353, arXiv:1810.12440, arXiv:1904.03493, arXiv:2010.04295, arXiv:1511.09207 |
| License | Gemma Terms of Use (gemma) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · PaliGemmaForConditionalGeneration · 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.
Weights-only memory for paligemma2-3b-pt-224: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 7.3 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 3.6 GB · Edge (<4 GB) |
| INT4 (0.5 bytes/param) | 1.8 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
finetune: 115quantized: 2adapter: 118
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.
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:40 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:12 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
