awesome-gemma
A curated directory of everything built on or around gemma-4 — models, runtimes, hosting, fine-tuning stacks, tutorials, demos, hackathon entries and papers, in nine sections. Small and new (30 stars, two commits on main when read on 2026-08-20), so its value here is not authority but coverage: it is the only inventory in the corpus that lists the whole Gemma surface at once.
Not an official Google product. The README's own footnote says so, and the repo is excluded
from Google's open-source vulnerability reward program — even though the google-gemma
org also hosts the Gemma cookbook the official docs link to. Treat the org name as
unresolved rather than as a Google byline. T3: real inclusion criteria and structure, but a
self-published directory that mostly relays vendor material. Most of what it points at is T1 —
ai.google.dev docs, model cards, the arXiv gemma-4-technical-report.
What it adds to this wiki
1. Gemma is a family of families. The catalog's largest section is not the core models but 16 named first-party variants — MedGemma, PaliGemma 2, ShieldGemma 2, TxGemma, VaultGemma, TranslateGemma and the rest. Paged separately as gemma-variants, because it changes what "gemma-4" names.
2. The 31B is real. gemma-4 carried the 31B-it as reported, not confirmed — the size
appeared in the corpus only as the base of a third-party derivative
(gemma-4-31b-it-scotoma), with no Google source naming it. The catalog links four
independent deployments of gemma-4-31b-it (Google's own AI Studio, OpenRouter, Cerebras,
an NVIDIA NVFP4 checkpoint) and cites a technical report "covering ... 31B". Confirmed at the
primary: the report's abstract gives the suite as 2.3B to 31B gemma-4-technical-report.
3. Fourteen hosted resellers, one open model. Beyond the local runtimes already recorded on
gemma-4, the Hosted section lists Vertex/Gemini Enterprise, OpenRouter, Cerebras, NVIDIA,
AMD, AI Studio, Cloud Run, LiveKit, Together AI, Modal, Fireworks, Baseten, Runpod and Cloudflare
Workers AI. Open weights do not remove the hosting market — they multiply it, which is the
distribution pattern open-weight-models describes and ai-gateway classifies. Six of the
fourteen have no page anywhere in the corpus; filed as a growth edge in synthesis.md.
4. Deployment reaches past the datacenter. Two orbital deployments are listed: Starcloud-1 ran Gemma on an H100 in orbit, and NASA runs it on satellite imagery, compressing pictures into text for low-bandwidth disaster response. Both are press links rather than first-party writeups, so they sit here as pointers, not findings.
What it does not do
No comparisons, no evaluations, no opinions — entries get a single descriptive line each. The one measurement it references is artificial-analysis on the 31B. Section headings like "Gemma 4 Good Challenge" and "Gemma in Space" are promotional framing carried over from Google's own campaigns. Read it as an index, and go to the primaries for anything load-bearing.
Related
gemma-4 · gemma-variants · gemma-4-technical-report · google · open-weight-models · gemma-4-qat · ai-gateway