Gemma 4 31B IT

Checkpoint identity

Official identifier

google/gemma-4-31B-itGoogle Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Architecture

Dense multimodal conditional-generation model; hybrid local sliding and global attentionGoogle Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Total parameters

Not verified for this recordPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Google Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Active parameters

Not verified for this recordActive parameters are a compute-related count, not the model's full stored weight count.Google Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Modalities

text-input, image-input, text-outputGoogle Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Context

262,144 tokensNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Google Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Release date

Exact release date not independently established in the reviewed artifactsGoogle Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Download access

Public checkpoint referenceGoogle Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

License

Apache 2.0Google documents Apache 2.0 for Gemma 4. These terms differ from the Gemma Terms of Use governing the existing Gemma 3 reference. Preserve applicable license and notice requirements for the exact checkpoint and derivative artifacts.Google Gemma4 31B IT official model card (external source) · Gemma4 31B IT checkpoint configuration (external source) · Google Gemma4 Apache2.0 license documentation (external source)

Theoretical weight-memory footprint

A weight estimate is unavailable without a verified parameter count.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

The main checkpoint configuration declares BF16. Quantized variants need their own artifact provenance and engine support; the 31B name does not specify a storage format or actual runtime-memory requirement.

Serving considerations

• The publisher's dense-model table reports 30.7B parameters and separately reports an approximately 550M vision encoder. The exact combined artifact boundary is left unknown in the single total field. It is a dense model, not the separate 26B-A4B mixture-of-experts variant.
• The text configuration confirms 262,144 positions. The 31B variant accepts text and images and has no audio encoder; audio capabilities of other Gemma 4 sizes must not be transferred to this checkpoint.
• The hybrid sliding/global attention config, image processing and visual tokens affect cache growth. A full-context estimate based solely on a conventional all-global dense cache should be treated as a preliminary model rather than a measured allocation.
• Google supplies an official model card and documentation. Engine compatibility remains unverified in this record; generic Hugging Face deployment snippets are not evidence of an executed or version-pinned Aguasto environment.
• The profile is a research reference. Aguasto has not tested its GPU fit, long-context quality, memory use or serving throughput and does not claim model availability.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.