Gemma 4 26B A4B IT

googlegoogle/gemma-4-26b-a4b-it

Google's Gemma 4 26B Mixture-of-Experts instruction-tuned model with 4B active parameters per token — open weights under Apache 2.0. Ranks #6 on the open Arena leaderboard. Multimodal text + image input. (Renamed from gemma-4-26b-moe to match official upstream ID gemma-4-26b-a4b-it.)

Gemma 4 model card ↗Gemma 4 model overview ↗Gemma 4 weights on Hugging Face ↗

Gemma 4 26B A4B IT is Google DeepMind's mixture-of-experts member of the Gemma 4 open-model family. The official Gemma 4 model card describes it as a 26B-total-parameter model with roughly 4B active parameters per token, text and image input, text output, thinking modes, native function calling, system-role support, and a context window up to 256K tokens. Google publishes Gemma 4 as open weights under Apache 2.0, with pre-trained and instruction-tuned variants across E2B, E4B, 12B, 26B A4B, and 31B sizes.

For TheRouter, Gemma 4 26B A4B IT is the efficient high-throughput Gemma 4 route when teams want multimodal understanding, agentic/function-calling coverage, and stronger benchmark headroom than the small E-series models without paying for a proprietary frontier model. It should not be treated as a Gemini replacement: the best fit is production chat, image-grounded support, document/PDF parsing, and coding or reasoning tasks where open-weight portability and predictable routing matter more than the absolute frontier ceiling.

Best for

• Vision-language extraction, inspection, and image-grounded support flows that need a 26B MoE open-weight model rather than a proprietary frontier model
• Multilingual assistants and content workflows where 140+ language coverage matters and model portability is a buying criterion
• Cost-sensitive document QA and extraction with long but bounded context, especially when 256K tokens is enough
• Teams evaluating open-weight deployment paths before deciding whether to self-host, fine-tune, or keep a managed API route

Reach for something else if

• Frontier reasoning, autonomous coding, or hard math where Gemini 2.5 Pro, Gemini 3-class models, Claude, or GPT-5.5-class models have a much higher ceiling
• Ultra-cheap routing, classification, or extraction where smaller Gemma, Llama, or Qwen models can meet quality targets at lower latency
• Native image, audio, or video generation — Gemma 4 26B A4B IT understands images but returns text, so use a dedicated generation model

How TheRouter serves this differently from the vendor

As the vendor operates it

Google publishes Gemma 4 26B A4B IT as open weights for self-hosting, fine-tuning, and deployment across local hardware, Hugging Face, Kaggle, Google AI Studio, Vertex AI, and other runtimes.

On TheRouter

TheRouter serves it as a managed OpenAI-compatible chat-completions route with text/image input and text output; this page does not claim local weight download, fine-tuning, or vendor-native runtime controls on the live TheRouter request path.

Google AI for Developers ↗

Context Length

128K

Max Output

Input Priceper 1M tokens

$0.162/ 1M tokens

Output Priceper 1M tokens

$0.648/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

Type	Rate
Input	$0.162 / 1M tokens
Output	$0.648 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_pstop

Specifications

Release date	2026-07-08Google AI for Developers ↗	verified
Architecture / size	Mixture-of-Experts transformer, 26B total parameters with 4B active parameters per tokenGoogle AI for Developers ↗	verified
Context window	Up to 256K tokensGoogle AI for Developers ↗	verified
Training tokens	Not publicly disclosed for the 26B A4B modelGoogle AI for Developers ↗	unknown
Input / output modalities	Text and images in; text outGoogle AI for Developers ↗	verified
License	Gemma Terms of Useai.google.dev ↗	verified

Benchmarks

Benchmark	Score	Source
MMLU Pro Gemma 4 26B A4B, 0-shot, as reported in Google's Gemma 4 model card	82.6%	ai.google.dev ↗
LiveCodeBench v6 Gemma 4 26B A4B, 0-shot, code generation benchmark from the official model card	77.1%	ai.google.dev ↗
AIME 2026 no tools Gemma 4 26B A4B, 0-shot, competition-math benchmark from the official model card	88.3%	ai.google.dev ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "google/gemma-4-26b-a4b-it",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-4-26b-a4b-it",
    "messages": [{"role": "user", "content": "Explain adaptive image tiling in two paragraphs."}]
  }'

More from google

google/gemini-2.5-flashManaged Google workhorse with stronger reasoning, longer context, and native Gemini tooling google/gemini-2.5-proGoogle's higher-ceiling reasoning option when quality matters more than open weights google/gemma-3-12bLarger Gemma 4 sibling when the workload needs more quality headroom while staying on open weights

Similar models

Cross-provider sibling models

mistral/ministral-3b

Compact open-weight alternative for teams comparing small multimodal routes and licensing tradeoffs

meta/llama-3.1-8b

Smaller open model for lower-latency text workloads without Gemma 4's vision focus

qwen/qwen3-14b

Alibaba's 14B-class open model alternative with strong multilingual and reasoning coverage

News & changes

2026-07-08

Google documents Gemma 4 with MoE, thinking, and 256K context

Google's launch post presents Gemma 4 as a five-size open-model family with E2B, E4B, 12B, 26B A4B, and 31B variants; the 26B A4B model is a mixture-of-experts option with roughly 4B active parameters per token, image-and-text input, text output, thinking, function calling, and up to 256K context.

re-authored by TheRouterGoogle AI for Developers ↗

Frequently asked

Is Gemma 4 26B A4B IT a good default model for production chat?

Use it when open weights, visual input, and predictable cost matter. For highest reasoning quality, longer context, or managed Google-native tools, choose a Gemini route instead.

Does Gemma 4 26B A4B IT generate images?

No. Gemma 4 26B A4B IT is a vision-language model that can read image input and return text. Use a dedicated image model for generation.

How is TheRouter serving different from Google's open-weight release?

Google's release lets teams download and run weights in multiple environments. TheRouter exposes a managed OpenAI-compatible API route, so integration is simpler but local runtime and fine-tuning controls are outside this route.

Are the API snippets verified?

No. This curator run has no paid API budget, so snippet verification remains unproven until a human operator approves and funds the verification run.

Fact ledger — every claim on this page traces here

source	URL	retrieved
Release date	Google AI for Developers ↗	2026-08-22	verified
Architecture / size	Google AI for Developers ↗	2026-08-22	verified
Context window	Google AI for Developers ↗	2026-08-22	verified
Training tokens	Google AI for Developers ↗	2026-08-22	unknown
Input / output modalities	Google AI for Developers ↗	2026-08-22	verified
License	ai.google.dev ↗	2026-08-22	verified
MMLU Pro	ai.google.dev ↗	2026-08-22	verified
LiveCodeBench v6	ai.google.dev ↗	2026-08-22	verified
AIME 2026 no tools	ai.google.dev ↗	2026-08-22	verified
Google documents Gemma 4 with MoE, thinking, and 256K context	Google AI for Developers ↗	2026-08-22	verified