Gemma 4 26B A4B is an instruction-tuned Mixture-of-Experts model from Google DeepMind. About 25B parameters total with roughly 3.8B active per token, so you get strong quality with lower inference cost than a dense model of similar capability. It supports multimodal input (text, images, and short video), function calling, and a 23K context window.
Gemma 4 26B A4B - NVIDIA RTX 4090
ollama/gemma4:26b
Gemma 4 26B A4B benchmarks, GPU requirements, and comparisonsgemma4:26bReleased April 2, 202623K context
Best forReasoningAgentic Coding
Best fit competitor·Claude Sonnet 4.5
Why teams pick Gemma 4 26B A4B over Claude Sonnet 4.5
Run it on your own GPU with predictable flat pricing — no per-token API meter running in the background.
Apache 2.0 weights you can fine-tune, audit, and keep inside your network instead of routing prompts through a hosted API.
MoE architecture activates only 3.8B parameters per token, so you get strong reasoning quality without paying for a full dense 27B+ API bill.
About
Gemma 4 26B A4B IT is an instruction-tuned Mixture-of-Experts (MoE) model from Google DeepMind. Despite 25.2B total parameters, only 3.8B activate per token during inference — delivering near-31B quality at a fraction of the compute cost. Supports multimodal input including text, images, and video (up to 60s at 1fps). Features a 256K token context window, native function calling, configurable thinking/reasoning mode, and structured output support. Released under Apache 2.0.
At a glance
Release
April 2, 2026
Parameters
25.8B (reported)
Quantization
Q4_K_M
Size
18GB
Context
23K
Benchmarks
Performance metrics for Gemma 4 26B A4B (Reasoning). Source: Artificial Analysis.
Performance indexes
31.2
Artificial Analysis
Intelligence Index
Better than 65% of models compared
22.4
Artificial Analysis
Coding Index
Better than 57% of models compared
32.1
Artificial Analysis
Agentic Index
Better than 58% of models compared
Benchmark scores
GPQA Diamond
iGraduate-level scientific reasoning
79.2%
HLE
iHumanity's Last Exam
18.3%
IFBench
iInstruction-following benchmark
72.4%
τ²-Bench Telecom
iConversational AI agents in dual-control scenarios
43.6%
AA-LCR
iLong context reasoning evaluation
55.7%
GDPval-AA
iEconomically valuable tasks
25.7%
CritPt
iResearch-level physics reasoning
0.0%
These figures are third-party or publisher-reported snapshots, not measurements from an OpenLLM Buddy GPU. Harness, dataset, and date belong to the original eval. Artificial Analysis methodology.
Apps & integrations
Choose an app below. Each guide shows how to point the app at your OpenAI-compatible endpoint.
n8n
Automate workflows and call your model as a node.
Open
OpenClaw
Build AI agents and tools on an OpenAI-compatible endpoint.
Open
Hermes
Connect agent runners to your chat completions endpoint.
Open
OpenCode
Power developer tools with your OpenAI-compatible model.
Open
Cursor
Override OpenAI Base URL in Cursor Settings and use your model with BYOK.
Open
VS Code
Use the Cline extension in VS Code to connect your OpenAI-compatible endpoint.
Open
Codex
Run OpenAI Codex CLI against your Chat Completions endpoint via config.toml.
Open
Raspberry Pi
Full Pi OS guide: SSH, API keys, curl, Python venv, systemd, and troubleshooting.
Open
Open WebUI
Self-hosted ChatGPT-style interface with your OpenAI-compatible endpoint as the backend.
Open
LibreChat
Multi-provider chat app. Add OpenLLM Buddy as a custom endpoint in librechat.yaml.
Open
SillyTavern
Roleplay and creative-writing front end. Use the Custom (OpenAI-compatible) chat source.
Open
Continue
Open-source AI assistant for VS Code and JetBrains. Add a model in config.yaml.
Open
Aider
Terminal pair programmer that edits your git repo. Point it at your endpoint with two flags.
Open
Zed
Fast native editor with a built-in AI agent. Add an OpenAI-compatible provider in settings.json.
Open
OpenAI SDK (Python & JS)
The official OpenAI client libraries, pointed at your endpoint by changing base_url.
Open
LangChain
Use ChatOpenAI with a custom base_url to build chains, tools, and agents on your model.
Open
LlamaIndex
Use OpenAILike to run RAG pipelines and agents against your OpenAI-compatible model.
Open
LiteLLM
One gateway for many providers. Route to your endpoint with the openai/ prefix, SDK or proxy.
Open
Dify
Build agents, chatbots, and workflows visually. Add OpenLLM Buddy as an OpenAI-compatible provider.
Open
Flowise
Drag-and-drop LLM flows. Use the ChatOpenAI Custom node with your base path.
Open
Langflow
Visual agent and RAG builder. Point the OpenAI component at your endpoint.
Open
Home Assistant
Run a private voice and chat assistant for your smart home on your own model.
Open
FAQ
Frequently asked questions
Common questions about Gemma 4 26B A4B, deployment, and using it on OpenLLM Buddy.
6 questions
Click Deploy on this page or open the console, pick a time pack, and launch an instance. When the pod is ready, you receive an OpenAI-compatible chat-completions URL and an API key scoped to your deployment.
Gemma 4 26B A4B is released under . You can run it on OpenLLM Buddy for production workloads; check the upstream model card for any additional attribution or usage notes from Google.
Use the OpenAI Chat Completions format against your instance URL. Set the model field to gemma4:26b in the JSON body—the same identifier shown at the top of this page. Official OpenAI SDKs work with a custom base URL and your API key.
You pay a flat rate per deployment pack (for example hourly or daily), not per token. See the Pricing tab for current pack prices. Token usage is metered for fairness and limits, but billing is tied to the pack you choose at deploy time.
Yes. The model supports configurable thinking/reasoning modes and native function calling where enabled in your request. Check the API tab for sample payloads and the Specs tab for capability details.
Ready to try it? Deploy Gemma 4 26B A4B · Browse AI models