DEV.co
Open-Source LLM · nvidia

diffusiongemma-26B-A4B-it-NVFP4

DiffusionGemma-26B-A4B-IT-NVFP4 is a quantized, multimodal LLM from NVIDIA based on Google's Gemma 4 architecture. It processes text, image, and video inputs to generate text output, with 25.2B total parameters but only 3.8B active due to its Mixture-of-Experts design. The NVFP4 quantization reduces model size and GPU memory requirements for faster inference. It supports 256K token context, function calling, and 35+ languages. Licensed under Apache 2.0 with Gemma terms overlay. Optimized for NVIDIA GPUs (Hopper, Blackwell) and served via vLLM.

Source: HuggingFace — huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4
14.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
1.7M
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developernvidia
Parameters14.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads1.7M
Likes95
Last updated2026-07-04
Sourcenvidia/diffusiongemma-26B-A4B-it-NVFP4

What diffusiongemma-26B-A4B-it-NVFP4 is

Transformer-based MoE architecture with bidirectional encoder-decoder attention and discrete diffusion sampling. Generates text in parallel 256-token blocks at >1,100 tokens/sec on H100 FP8. Input: text, RGB images (variable resolution via 70–1120 token budgets), video (up to 60 sec @ 1 fps). Output: text with native function calling and structured JSON. NVFP4 quantization reduces FP16→4-bit activations/weights. Inference via vLLM with TRITON_ATTN backend. Evaluation on GPQA Diamond, AIME 2025, GSM8K, HumanEval, MMLU shows <1% accuracy loss vs. BF16 baseline. Training data cutoff: January 2025; includes CSAM filtering and PII removal.

Quickstart

Run diffusiongemma-26B-A4B-it-NVFP4 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="nvidia/diffusiongemma-26B-A4B-it-NVFP4")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Conversational AI & Chatbots

Supports multi-turn dialogue with thinking mode enabled for step-by-step reasoning and function calling for tool integration. Compact footprint post-quantization suitable for cost-effective deployment at scale.

Multimodal Document & Image Understanding

Process OCR, PDF parsing, chart comprehension, UI/screenshot analysis, and video content analysis in a single model. Variable image token budgets (70–1120) allow cost-accuracy tradeoffs per workload.

Code Generation & Reasoning Tasks

HumanEval 95.0% (NVFP4) indicates strong coding capability. Thinking mode supports complex step-by-step logic. MoE design keeps inference cost low despite 25.2B parameters.

Running & fine-tuning it

ESTIMATE: 4-bit quantization reduces memory significantly from FP16. Approximate VRAM: 14–26 GB for inference on single H100/B100 GPU (quantized weights ~13 GB + activation buffers). Inference tested on NVIDIA Blackwell B100; compatible with Hopper H100. Requires CUDA-capable NVIDIA GPU; Linux preferred. Multi-GPU setups supported via vLLM for higher concurrency. Exact VRAM/throughput depends on batch size, context length, and image/video input token budgets.

Not stated in card. Post-training quantization (NVFP4) completed by NVIDIA ModelOpt. No mention of LoRA, QLoRA, or fine-tuning feasibility on quantized weights. Requires review of vLLM and ModelOpt documentation or direct testing. Base model (Gemma 4) fine-tuning methodology unknown; assume standard transformer SFT/DPO may apply, but on quantized weights feasibility is unclear.

When to avoid it — and what to weigh

  • CPU-only deployments — Model is explicitly optimized for NVIDIA GPUs (Hopper, Blackwell). CPU inference is not a supported deployment path; expect severe latency penalties.
  • Real-time safety-critical applications without additional testing — Card notes model may amplify toxic language and societal biases from training data. Requires iterative testing and validation per V-model methodology before production deployment in safety-sensitive domains.
  • Low-latency, single-digit token generation — Diffusion sampling generates 256-token blocks in parallel, not ideal for latency-sensitive streaming use cases requiring immediate output of first tokens.
  • Proprietary/closed-source model modifications without Apache 2.0 compliance — Apache 2.0 license requires attribution and disclosure; Gemma Terms of Use overlay adds restrictions. Closed-source derivative products require legal review.

License & commercial use

Apache License 2.0 (SPDX-License-Identifier: Apache-2.0) with two additional terms: (1) Google Gemma Terms of Use, (2) Google Gemma Prohibited Use Policy. Apache 2.0 is OSI-approved and permissive, requiring attribution and liability disclaimer. Gemma overlay restricts use to non-commercial research and educational purposes unless explicit commercial agreement signed with Google. NVIDIA model card states 'ready for commercial and non-commercial use', but this conflicts with Gemma TOU if Google's commercial license is not independently secured.

Requires review. The model card claims 'ready for commercial and non-commercial use', but is governed by Apache 2.0 + Google Gemma Terms of Use + Gemma Prohibited Use Policy. Apache 2.0 alone permits commercial use, but Gemma TOU restricts commercial deployment unless you have a separate commercial agreement with Google. NVIDIA disclaimer notes this is a third-party (Google) model quantized by NVIDIA. Before deploying commercially, verify: (1) Direct Apache 2.0 commercial eligibility is asserted only by NVIDIA quantization layer, not by Google's Gemma base model. (2) Google's commercial Gemma license status. Consult legal before production commercial use.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityNeeds review
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

Card does not claim the model 'is secure'. Security notes: (1) Training data includes CSAM filtering and PII removal—mitigates some data-sourced risks but no guarantee of zero leakage. (2) Model trained on internet-scale data containing toxic language and societal biases; may amplify these when prompted. (3) No mention of prompt injection, jailbreak testing, or adversarial robustness. (4) Native function calling feature expands attack surface if downstream tools are called without sandboxing. (5) Gemma Prohibited Use Policy restricts illegal, deceptive, and harmful uses; enforcement is policy-dependent, not technical. Conduct threat modeling and testing before production deployment, especially for safety-sensitive applications.

Alternatives to consider

Llama 3.1 70B (Meta)

Open-weights, OSI-licensed (Llama Community License), larger context (128K), broader tool ecosystem. Quantized variants available. Slightly heavier footprint but established production deployments and stronger documentation.

Mixtral 8x22B (Mistral)

Similar MoE efficiency (8 experts, ~13B active). Apache 2.0 license without Google overlay. Smaller total parameters (141B vs. 25.2B) but finer control. Strong vLLM support.

Claude 3.5 Sonnet (Anthropic, API-only)

Proprietary, closed-source, hosted only. No quantization needed. Strong multimodal (image/video) support and reasoning. Trade-off: no self-hosting, vendor lock-in, higher per-token cost but production-ready safety.

Software development agency

Ship diffusiongemma-26B-A4B-it-NVFP4 with senior software developers

This model demands NVIDIA GPU infrastructure (Hopper H100 or Blackwell B100) and vLLM orchestration. Before production, verify: Gemma commercial licensing with Google, benchmark VRAM/latency on your hardware, and test for toxic output mitigation in your use case. Contact our AI engineering team to validate deployment architecture and compliance.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

diffusiongemma-26B-A4B-it-NVFP4 FAQ

Can I use this model for commercial applications?
The NVIDIA model card states 'ready for commercial and non-commercial use', but the model is governed by Apache 2.0 + Google Gemma Terms of Use. Apache 2.0 permits commercial use with attribution. However, Gemma TOU may impose restrictions unless you have a separate commercial license from Google. Before deploying to production for commercial purposes, consult legal and verify Google's commercial Gemma licensing status.
What is the estimated VRAM requirement for inference?
Approximate estimate: 14–26 GB VRAM on a single H100 or B100 GPU for batch size 1–4 at full context. The NVFP4 4-bit quantization significantly reduces memory vs. FP16 (~52 GB). Exact VRAM depends on batch size, context length, and image/video token budgets. Test on your hardware before production deployment.
How does inference speed compare to the full-precision baseline?
Card states >1,100 tokens/sec on H100 FP8. NVFP4 quantization maintains similar throughput with lower memory and GPU footprint. Exact speedup over FP16 baseline is not disclosed; vLLM + TRITON_ATTN backend is primary optimization. Benchmark on your inference setup for definitive numbers.
Is fine-tuning supported on the NVFP4 quantized weights?
Not clearly stated. Post-training quantization is complete; no mention of LoRA, QLoRA, or fine-tuning workflows on quantized weights. Assume fine-tuning requires reverting to full-precision or base model. Requires direct testing or consultation with NVIDIA ModelOpt and vLLM documentation.

Software development & web development with DEV.co

Need help beyond evaluating diffusiongemma-26B-A4B-it-NVFP4? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy DiffusionGemma? Assess Your Infrastructure.

This model demands NVIDIA GPU infrastructure (Hopper H100 or Blackwell B100) and vLLM orchestration. Before production, verify: Gemma commercial licensing with Google, benchmark VRAM/latency on your hardware, and test for toxic output mitigation in your use case. Contact our AI engineering team to validate deployment architecture and compliance.