DiffusionGemma 26B A4B / checkpoint worksheet

48.10 GiB at BF16. 17.53 GiB at NVFP4.

Q: How much GPU memory does DiffusionGemma need?

Google's official BF16 checkpoint contains 11 safetensors files totaling 51,647,701,024 bytes, or 48.10 GiB. NVIDIA's official NVFP4 checkpoint contains two safetensors files totaling 18,823,855,888 bytes, or 17.53 GiB. Both require additional runtime and context memory.

Q: Can DiffusionGemma run on a 24 GB GPU?

The 17.53 GiB NVIDIA NVFP4 files plus a generic 20 percent reserve total about 21.04 GiB. That clears 90 percent of a 24 GiB device by only about 0.56 GiB before context state, so it is a tight arithmetic fit rather than a supported deployment claim. NVIDIA names the 32 GB RTX 5090 for prototyping.

Q: Can DiffusionGemma BF16 run on one 48 GB GPU?

No straightforward full-GPU plan fits the official 48.10 GiB checkpoint on a nominal 48 GiB device before runtime workspace or cache. With a generic 20 percent reserve, the target is about 57.72 GiB, so an 80 GiB-class card provides a more credible single-device planning floor.

Q: Does DiffusionGemma only store its 3.8B active parameters?

No. DiffusionGemma has 25.2B total parameters and 3.8B active parameters. Sparse activation reduces per-token compute, but the full expert checkpoint still needs storage and routing.

Q: How does 256K context affect DiffusionGemma GPU memory?

The official model card lists context up to 256K tokens. DiffusionGemma uses an encoder-decoder block-diffusion design with an encoder KV cache and bidirectional decoder canvases, so context memory should be measured in the selected runtime rather than inferred from a generic dense decoder-only formula.

DiffusionGemma activates 3.8B of 25.2B parameters, but the full expert checkpoint still has to reside somewhere. Start with the exact files, then budget the encoder cache, decoder canvases, runtime workspace, and concurrency.

Open the editable RTX 5090 NVFP4 plan →

Quick answer / artifact floors

A 24 GB NVFP4 fit is tight. A 48 GB BF16 fit is short.

These plans use exact official artifact totals and a visible generic reserve. They do not claim that every kernel, context length, or runtime supports the resulting device.

NVFP4 · 24 GiB card 0.56 GiB margin Artifact + 20% reserve vs 90% usable; cache excluded → NVFP4 · RTX 5090 7.76 GiB margin NVIDIA-named prototyping hardware; measure context → BF16 · 48 GiB card SHORT 0.10 GiB Files alone exceed the nominal device before runtime → BF16 · 80 GiB card 57.72 GiB target Exact files + 20% reserve; 256K state excluded →

Official files vs runtime

The quantized checkpoint changes the floor, not the architecture.

Google's model card defines the architecture and BF16 checkpoint. NVIDIA's model card defines its optimized NVFP4 release and supported format. The arithmetic below keeps file fit separate from serving fit.

Swipe the matrix horizontally →

Input	Official value	What it changes	What it does not prove
Google BF16 checkpoint	48.10 GiB · 11 files	Resident-weight floor	Runtime or 256K capacity
NVIDIA NVFP4 checkpoint	17.53 GiB · 2 files	Quantized artifact floor	Kernel support on every GPU
Total / active	25.2B / 3.8B parameters	Storage vs per-step compute	3.8B-size memory usage
Generation	256-token diffusion canvases	Parallel token refinement	Decoder-only KV-cache behavior
Context	Up to 256K tokens	Maximum advertised prompt window	Full-window concurrency
NVIDIA prototyping	RTX 5090 or DGX Spark	Named first-party starting points	A universal minimum GPU

Artifact totals were calculated from the official Google and NVIDIA Hugging Face repositories on August 7, 2026. Recheck the inventories after checkpoint revisions.

Read the estimate

Active is not resident.
3.8B active parameters reduce compute, but all 25.2B weights remain part of the checkpoint.

NVFP4 is format plus hardware.
A small artifact does not guarantee compatible kernels or good throughput on every GPU.

256K is runtime state.
The encoder cache and decoder canvases must be measured under the chosen concurrency.

Context and cache boundary

Diffusion generation changes which cache math is safe.

Official architecture
25.2B total · 3.8B active · 30 layers

Generation unit
256-token canvas · bidirectional attention

Context
up to 256K · encoder KV cache

The calculator links hold context at zero because a generic decoder-only cache formula would create false precision. Profile the selected Transformers, NeMo AutoModel, NIM, or other supported runtime at the intended prompt length and batch size.

Primary-source inputs

Start with Google and NVIDIA's exact releases.

Architecture, context, and the BF16 files come from Google's model card and repository. NVFP4 file size and supported format come from NVIDIA's official optimized checkpoint and technical release.

Open the GPU memory calculator → Compare the Gemma 4 family → Find models for your GPU budget → Official Google model card → Official Google BF16 checkpoint → Official NVIDIA NVFP4 checkpoint → NVIDIA deployment overview →

How much GPU memory does DiffusionGemma need?

The Google BF16 files total 48.10 GiB. NVIDIA's NVFP4 files total 17.53 GiB. Both need additional runtime and context memory.

Can DiffusionGemma run on a 24 GB GPU?

The NVFP4 artifact plus a generic 20% reserve leaves only about 0.56 GiB inside a 90%-usable 24 GiB plan before context state. Treat that as tight arithmetic, not a supported topology.

Can DiffusionGemma BF16 run on one 48 GB GPU?

The 48.10 GiB files already exceed a nominal 48 GiB device. A generic 20% reserve moves the planning target to 57.72 GiB before context state.

Does DiffusionGemma only store its 3.8B active parameters?

No. The model has 25.2B total parameters. Sparse activation reduces per-step compute, not the storage required for the full expert checkpoint.

How does 256K context affect DiffusionGemma GPU memory?

The model uses an encoder KV cache and bidirectional diffusion canvases. Measure full-window memory in the chosen runtime instead of applying a generic decoder-only formula.