2 Apr 2026
gemma-4-31B-it
The dense control for Gemma’s MoE experiment
30.7B reported total · text + image + video
Gemma / MoE
gemma-4-26B-A4B-it
Sparse compute in a workstation-sized family
New to model sizes or GPU memory? Start with weights, parameters and quantization →
This model is a useful MoE comparison for the dense Gemma 4 31B. Evaluate both with the same visual inputs and response budgets; the A4B label should lead to a compute discussion, not a four-billion-weight memory estimate.
25.2Breported total parameters
3.8Bactive parameters per token
262,144configured token positions
Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Evaluate it for
Mid-sized text, image and video deployments and dense-versus-MoE studies.
Choose another path when
Audio input inferred from other Gemma variants, or capacity planning based on the active count.
Release context
Why this release matters
The April family offered both dense and MoE choices. The reported 25.2B total and 3.8B active values make the rounded 26B-A4B name precise enough to reason about.
Deployment starting points
Which hardware, how many, at what precision?
Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.
No certified hardware configuration in this edition
Audio input inferred from other Gemma variants, or capacity planning based on the active count.
The pinned model card establishes the architecture. It does not, by itself, prove that a particular GPU count runs this checkpoint at your target context. The capacity screen below helps rule out undersized allocations; a successful load and workload test are still required.
51.61 GB of tensor data in the pinned index
This is the index’s reported tensor payload, in decimal GB. It excludes file headers and runtime memory. Mixed precision, conversion and host/device placement determine how much GPU memory the loaded model needs.
Inspect the exact index metadata →Capacity screen: would the weights fit?
This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.
| Representation | Approximate weight floor | What remains to budget |
|---|---|---|
| 16-bit uniform | 50.4 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 8-bit uniform | 25.2 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 4-bit uniform | 12.6 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
Enter the total across the whole model replica, not per GPU: KV or recurrent cache for all concurrent requests, image/audio state and execution workspace. There is no universal default. We additionally leave 15% of each GPU unused.
50.4 GB for hypothetical weights
Enter a request-and-buffer budget to calculate the aggregate capacity requirement.
A count of one does not prove the engine supports this GPU. A count above one does not prove the model shards evenly or communicates efficiently. Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Architecture in practice
What the serving engine has to do
Local and global layers read different spans.
Window length and layer ratio are checkpoint-specific.
Full-context state remains for global layers.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
Inspect the full checkpoint specifications
| Exact checkpoint | google/gemma-4-26B-A4B-it |
|---|---|
| Text model type | gemma4_text |
| Layers | 30 |
| Hidden width | 2,816 |
| Attention / KV heads | 16 / 8 |
| Head dimension | 256 |
| Routed / selected experts | 128 / Not verified |
| Configured positions | 262,144 |
| Documented extension | Not recorded for this checkpoint |
| Inputs → output | text, image, video → text |
| License metadata | apache-2.0 |
25 sliding attention layers; 5 full attention layers. These counts describe the configured pattern.
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
What else belongs on the shortlist?
The dense sibling isolates feed-forward sparsity; Qwen provides a different hybrid-attention MoE.
24 Feb 2026
Qwen3.5-35B-A3B
The easiest MoE memory lesson to make concrete
35B reported total · text + image + video
These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.
Prepare to explain it
The interview lesson
Explain sliding-window versus full-attention state and why the same cache formula cannot simply be multiplied across all layers.
Attention layouts · Quantization · Courses and worked examples
For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.
Sources and citation
Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.
AI Infra Interviews, “gemma-4-26B-A4B-it: release, architecture and deployment”, checked 2026-09-13. Preserve this date and the exact checkpoint when citing.