Consider it when…
- Test image-and-text question answering with bounded inputs
- Prototype multimodal document or screenshot interpretation
- Compare compact local vision-language behaviour before scaling up
google/gemma-3-4b-itA compact Gemma instruction model that accepts text and images and returns text for local multimodal experiments.
A compact multimodal instruction model with gated model files.
The pinned publisher record classifies this release as image-text-to-text with gated Gemma files.
google/gemma-3-4b-it093f9f388b31These are Maviter calculations at an 8,192-token context—not publisher hardware requirements or proof that a runtime supports this release.
Parameters × 4 bits, converted to GiB.
Architecture evidence is incomplete in the current record.
Maviter will not invent a complete estimate.
Inspect the architecture and exact serving runtime.
Run the exact publisher release, on the intended runtime, with inputs from the work it must actually perform.
093f9f388b31de276ce2de164bdc2081324b9767Open source ↗Pinned model cardCapabilities, licence and usage notesREADME.md at the recorded revisionRead card ↗The evidence ledger pins an exact publisher revision and monitors it for drift. A source check does not independently reproduce benchmarks or certify production suitability.
The publisher provides discoverable weights, but runtime support and performance are machine-specific. Use Maviter's fit calculator, then run one controlled test.
This record documents downloadable weights. It does not assert a provider-hosted API ID, current price, rate limit or data policy.
No. The pinned links return to the publisher. Maviter does not mirror binaries or endorse unaudited third-party quantisations.