DEV.co
Open-Source LLM · nvidia

Gemma-4-26B-A4B-NVFP4

Gemma-4-26B-A4B-NVFP4 is NVIDIA's quantized version of Google's Gemma 4 multimodal LLM. It uses a mixture-of-experts architecture with 3.8B active parameters out of 25.2B total, optimized for inference on NVIDIA GPUs via vLLM. The model supports text, images, and video inputs with a 256K token context window and handles reasoning, coding, summarization, and conversational tasks. It is released under Apache 2.0 license and marked as ready for commercial and non-commercial use. The NVFP4 quantization reduces model size while maintaining near-baseline performance on reasoning and coding benchmarks.

Source: HuggingFace — huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4
14.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
2.2M
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
Downloads2.2M
Likes112
Last updated2026-05-11
Sourcenvidia/Gemma-4-26B-A4B-NVFP4

What Gemma-4-26B-A4B-NVFP4 is

Gemma 4 26B A4B is a sparse mixture-of-experts transformer with 30 layers, 262K vocabulary, and 8 active experts out of 128 total. It employs hybrid attention (local sliding-window at 1024 tokens + global attention), proportional RoPE for long-context support, and unified KV in global layers. Vision encoder adds ~550M parameters for image/video processing. Quantized to NVFP4 (4-bit custom NVIDIA format) using Model Optimizer v0.43.0. Calibrated on cnn_dailymail and Nemotron datasets. Requires vLLM with tensor parallelism=1 (expert parallelism and Flashinfer support pending upstream fixes). Tested on NVIDIA B200 hardware. Training data cutoff January 2025, covering 140+ languages.

Quickstart

Run Gemma-4-26B-A4B-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/Gemma-4-26B-A4B-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

High-performance inference on NVIDIA GPUs with modest footprint

The sparse MoE design and NVFP4 quantization allow near-full-model performance with ~50% memory savings versus full-precision baselines. Ideal for cost-efficient production inference on NVIDIA Blackwell or newer architectures where vLLM is available.

Multimodal code-and-reasoning applications

Strong performance on AIME 2025 (90%), LiveCodeBench (79.8%), and IFBench (78.1%) makes it suitable for coding assistants, technical support bots, and agentic workflows that combine text, images, and structured outputs.

Multilingual conversational and summarization workloads

Trained on 140+ languages with a 256K context window, applicable to global customer support, long-document analysis, and cross-lingual RAG or chat systems. Apache 2.0 license enables open-source product integration.

Running & fine-tuning it

ESTIMATE: ~14–16 GB VRAM for NVFP4 quantized inference (4-bit weights + activations + KV cache for ~256K context). Full-precision baseline (bf16) would require ~45–50 GB. Verified on NVIDIA B200; should support L40S, H100, L100, and newer Blackwell GPUs. vLLM serving typically batches efficiently but does not yet support multi-GPU expert parallelism. Single-GPU deployments recommended until vLLM EP and Flashinfer support is released. Linux OS required.

Not clearly stated in model card. Base Gemma 4 26B A4B IT is a sparse MoE model; fine-tuning a quantized NVFP4 model is uncommon in practice. Recommend: (1) fine-tune the unquantized baseline (google/gemma-4-26B-A4B-it) with LoRA/QLoRA on NVIDIA hardware, then quantize with ModelOptimizer afterward; (2) or apply LoRA adapters in inference without re-quantizing. Requires review of NVIDIA ModelOptimizer and vLLM LoRA support compatibility. Sparse MoE fine-tuning may require specialized tooling (e.g., handling expert routing updates).

When to avoid it — and what to weigh

  • Deployment outside NVIDIA GPU ecosystems — Model is explicitly designed and optimized for NVIDIA hardware (CUDA, TensorRT, vLLM). CPU or non-NVIDIA accelerator deployment is unsupported and will see significant performance degradation. Consider CPU-optimized models (e.g., llama.cpp compatible) if heterogeneous infrastructure is required.
  • Distributed inference requiring expert or tensor parallelism across multiple GPUs — Current vLLM support is limited to TP=1 with pending issues in expert parallelism and Flashinfer backend integration. Large-scale distributed serving requires waiting for upstream vLLM PRs and releases, or using non-quantized baseline with TensorRT-LLM.
  • Production systems requiring deterministic or bias-controlled outputs — Model card explicitly states the base model was trained on toxic internet-crawled data and may amplify societal biases or produce toxic responses. Requires internal alignment testing, prompt engineering, or fine-tuning before use in sensitive domains (healthcare, legal, finance).
  • Use cases requiring video as native input (not frame-by-frame sequences) — Video support is noted on E2B/E4B model variants, not this 26B A4B model. Video must be decomposed into frames (up to 60 seconds at 1 FPS). If native video understanding is critical, verify suitability or consider alternative architectures.

License & commercial use

Apache License 2.0 (OSI-approved permissive license). Covers both the quantized NVIDIA model artifact and the underlying Google Gemma 4 26B IT base model. Allows modification, distribution, and private use with minimal restrictions (attribution required). See https://ai.google.dev/gemma/apache_2 for full terms.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use. Model card states 'This model is ready for commercial/non-commercial use.' No gating or commercial license required. However, developers must ensure: (1) compliance with any third-party data rights in training data (model trained on web-crawled content); (2) internal review for bias and safety in regulated domains; (3) proper input validation for images/video (personal data, IP, CSAM filtering per model card note). Recommend legal review if deploying in healthcare, financial services, or other high-risk sectors.

DEV.co evaluation signals

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

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

Model card does not claim or audit security posture. Considerations: (1) Base model trained on web-crawled data with filtering for CSAM and sensitive data, but comprehensive security audit not detailed; (2) NVFP4 quantization does not inherently introduce or mitigate vulnerabilities but reduces model size (smaller attack surface for exfiltration); (3) vLLM inference server requires standard hardening (auth, TLS, input validation); (4) Model supports function-calling and tool-use (--enable-auto-tool-choice flag) — ensure tool definitions are validated and sandboxed; (5) Multimodal inputs (images, video) require validation to prevent prompt injection or adversarial perturbations. No formal threat model or pen-test data provided. Recommend standard LLM security practices: input filtering, rate limiting, output filtering, audit logging.

Alternatives to consider

google/gemma-4-26B-A4B-it (unquantized baseline)

Full-precision original model; better for fine-tuning, research, or if VRAM availability allows. No quantization artifacts; wider tool/library support. Trade-off: ~3x larger and slower on consumer GPUs.

meta-llama/Llama-3.1-405B-Instruct (quantized variants)

Larger dense model (405B params) with strong reasoning/coding performance and broader industry adoption. Suitable if deployment budget allows multi-GPU clusters or if you need stronger inference support across vLLM, TGI, and llama.cpp. Trade-off: no official MoE sparsity; higher latency per token on single GPU.

mistralai/Mistral-Large-v2 (smaller quantized alternatives)

Smaller dense alternatives (e.g., 34B) with broader quantization framework support (GPTQ, AWQ, bfloat16 via Ollama). Better portability if NVIDIA-specific optimization is not a hard requirement. Trade-off: lower reasoning/coding performance than Gemma 4.

Software development agency

Ship Gemma-4-26B-A4B-NVFP4 with senior software developers

Verify NVIDIA GPU availability, set up vLLM with the provided command, and test on your target hardware before scaling. Consult our security and fine-tuning guides for regulated use cases.

Talk to DEV.co

Related open-source tools

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

Gemma-4-26B-A4B-NVFP4 FAQ

Can I use this model commercially without a license agreement?
Yes. Apache 2.0 is a permissive OSI license that permits commercial use without a separate commercial license. However, you must: (1) include a copy of the Apache 2.0 license; (2) state significant changes; (3) review third-party data rights (model trained on web-crawled data); (4) internally validate for bias/safety in regulated industries. Recommend legal review if deploying in healthcare, financial, or other high-risk sectors.
What GPU do I need to run this model?
NVIDIA Blackwell (B200) or newer preferred. Estimated ~14–16 GB VRAM for NVFP4 quantized inference. Also supported on H100, L40S, L100 GPUs with CUDA support. Linux OS required. CPU and non-NVIDIA accelerators are unsupported. Single-GPU inference only (TP=1); multi-GPU expert parallelism pending vLLM updates.
Can I fine-tune this quantized model?
Fine-tuning the NVFP4 quantized artifact directly is not standard practice. Recommended approach: fine-tune the unquantized base model (google/gemma-4-26B-A4B-it) with LoRA/QLoRA, then re-quantize with NVIDIA Model Optimizer. Alternatively, apply LoRA adapters at inference time. Requires review of vLLM LoRA support and sparse MoE fine-tuning tooling.
What are the known limitations of the vLLM implementation?
Current vLLM support (v0.20.0+) has: (1) Tensor parallelism only (TP=1); expert parallelism (EP) pending (GitHub issue vllm#39595); (2) MoE backend limited to VLLM_CUTLASS or Marlin; Flashinfer-TRT backend pending (PR vllm#41050); (3) No explicit throughput or latency benchmarks provided. Verify upstream vLLM roadmap for resolution timelines before production deployment.

Work with a software development agency

Need help beyond evaluating Gemma-4-26B-A4B-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 Gemma 4 in Production?

Verify NVIDIA GPU availability, set up vLLM with the provided command, and test on your target hardware before scaling. Consult our security and fine-tuning guides for regulated use cases.