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...
Gemma 4 26B A4B is a Mixture-of-Experts model developed by Google. It has 26 billion total parameters but only 4 billion are active per token—this design reduces computational cost while aiming to…
Gemma 4 26B A4B accepts text, image, and video as input. Images can be provided as base64-encoded data or URLs. Video can be provided as a URL or as a sequence of frames (image objects). The model processes these modalities jointly, enabling tasks like visual question answering, video summarization, and diagram understanding. Audio is not supported; only visual and textual content. The output is text only. The model’s multimodal capability is particularly useful for analyzing documents that contain charts, screenshots, or video recordings.
The context window is 262,144 tokens. This allows the model to process very long sequences in a single pass—for example, a 200-page document, hours of transcribed video, or a large set of images with descriptive captions. Larger context windows reduce the need for chunking and summarization, but they also increase memory usage. The effective length you can use will depend on the total number of input tokens (text + image/video tokens). Be mindful that image and video inputs consume many tokens; refer to OrcaRouter's documentation for how token counts are calculated for non-text inputs.
If your task is purely text-based, requires only a short context (under 8k tokens), or does not need multimodal input, consider a smaller or cheaper model—such as Gemma 3 4B or a text-only variant. Gemma 4 26B A4B is priced at $0.06 per million input tokens and $0.33 per million output tokens. For simple question answering or classification, models with lower per-token costs may be more economical. The MoE design makes it efficient relative to its total size, but not the cheapest option available on OrcaRouter for minimal tasks.
GPQA Diamond is a benchmark of 448 graduate-level multiple-choice questions across biology, physics, and chemistry. A score of 79.2 means the model answered 79.2% correctly. This indicates strong scientific reasoning and knowledge retrieval. The benchmark is designed to be difficult for many LLMs. However, a single benchmark cannot capture all aspects of model quality. For example, the model’s performance on other tasks like coding or creative writing may differ. Use this score as one data point when comparing models for similar scientific reasoning tasks.
Strengths include multimodal understanding with large context, MoE efficiency for its size, and strong reasoning on scientific questions as indicated by GPQA. Limitations are not exhaustively documented but are typical of MoE models: performance can vary by domain, and the effective capacity per token is limited by the 4B active parameters. The model may struggle with tasks requiring extremely deep logical chains or domain-specific jargon not well represented in training data. Latency and throughput depend on the deployment hardware; OrcaRouter does not guarantee specific speed metrics.
OrcaRouter does not publish standardised latency benchmarks for this model. As an MoE model, Gemma 4 26B A4B activates only a subset of parameters per token, which can make inference faster than a dense 26B model but possibly slower than a smaller dense model. Actual performance depends on factors such as batch size, input length, and backend GPU type. For real-time applications, test with your specific workload. You may also consider the trade-off between latency and cost—using a smaller model may improve speed at lower cost.
Pricing is $0.06 per 1 million input tokens and $0.33 per 1 million output tokens. These are the rates billed by the provider (Google) and passed through by OrcaRouter with zero markup. That means you pay exactly the provider rate—OrcaRouter does not add any surcharge. Tokens are counted consistently across the platform; images and video frames are tokenised according to Google’s model specifications. For a typical multimodal query with a few images, input tokens may dominate, making input pricing the main cost driver.
OrcaRouter may offer caching mechanisms for repeated prefix or prompt templates, which can reduce token consumption and lower costs. However, specific caching discounts are not guaranteed for this model and depend on your usage pattern. There is no separate batch pricing tier published for Gemma 4 26B A4B. For high-volume workloads, contact OrcaRouter support to discuss potential discounts. As with all models on the platform, you are only charged for what you use—input and output tokens—with no monthly fee or minimum commitment.
Given the pricing structure, the total cost depends on the number and type of tokens you send and receive. Multimodal inputs (especially video) can use many input tokens because each frame is encoded. For long videos, the input cost may exceed the output cost. If your task is output-heavy (e.g., generating long reports), the output price ($0.33/M) is higher than input. Evaluate your expected token ratio. For tasks that can be solved with a cheaper text-only model, the cost difference can be significant. Use OrcaRouter’s token counting tools to estimate.
Set the base URL to https://api.orcarouter.ai/v1 and use the model ID google/gemma-4-26b-a4b-it. Send a POST request to /chat/completions with the standard OpenAI schema. For multimodal input, include an array of content objects with the type field set to 'text', 'image_url', or 'video_url'. Example: messages: [{ role: 'user', content: [{ type: 'text', text: 'Describe this video.' }, { type: 'video_url', video_url: { url: 'https://example.com/video.mp4' } }] }]. The API will return a chat completion response.
You can use standard OpenAI parameters like temperature, top_p, max_tokens, stop, frequency_penalty, and presence_penalty. Additionally, OrcaRouter supports provider-specific parameters through the optional 'provider' field in the request body (not required for this model). The model natively supports streaming by setting stream=true. For structured outputs, use the 'response_format' parameter with type 'json_object' or a JSON schema. Refer to OrcaRouter’s documentation for any extra parameters like 'reasoning_effort'—though that is not listed for this model.
Switching from another OpenAI-compatible API is straightforward: change the base URL to https://api.orcarouter.ai/v1 and set the model to google/gemma-4-26b-a4b-it. Your existing prompt structure, parameters, and SDK client are compatible because OrcaRouter follows the same schema. If you were using a different provider’s SDK, you may need to update the endpoint and authentication. OrcaRouter uses API keys rather than OAuth; include your key in the Authorization header as 'Bearer YOUR_KEY'. Test with a small request first.
Gemma 4 26B A4B is a newer multimodal MoE model with 262k context and a GPQA Diamond score of 79.2, while Gemma 3 8B is a smaller dense model (8B parameters) with a context window of 128k and no native video support. Gemma 3 8B is cheaper on input tokens (typically $0.05-0.10 per million input) but may not match reasoning quality on hard scientific questions. For tasks involving video or very long documents, Gemma 4 26B A4B is the clear choice. For text-only tasks with moderate context, Gemma 3 8B may be sufficient and more cost-effective.
Llama 3.1 70B is a dense model with 70B parameters and 128k context, not natively multimodal for video (though it can process images). Gemma 4 26B A4B uses MoE to activate only 4B parameters per token, potentially offering faster inference than the much larger Llama model. On GPQA Diamond, Gemma 4 26B A4B scores 79.2; Llama 3.1 70B scores around 65-70 (not directly comparable due to benchmark version differences). Llama 3.1 70B may be more expensive on input tokens (around $0.35 per million input). For multimodal and long-context scenarios, Gemma 4 may be more efficient.
GPT-4o is a dense proprietary model from OpenAI with multimodal support and a context window of 128k (standard) and up to 1M for some versions. Its pricing is significantly higher (e.g., $2.50 per million input tokens for GPT-4o). Gemma 4 26B A4B is open-weight and available through OrcaRouter at a much lower cost ($0.06/$0.33). Performance on GPQA Diamond for GPT-4o is not directly comparable, but typically higher. However, for cost-sensitive applications that do not require frontier-level reasoning, Gemma 4 26B A4B offers a strong price-performance ratio. Data handling differs: Gemma 4 is from Google with separate privacy terms.
Compared to other open-weight MoE models like Mixtral 8x7B (46.7B total, 12.9B active) or Qwen2.5-72B-A3B (72B total, 3B active), Gemma 4 26B A4B offers a unique combination: a 262k context window, full multimodal support (image+video), and a published GPQA Diamond score of 79.2. Mixtral 8x7B has a 32k context and no video support. Qwen2.5-72B-A3B has a 128k context and supports text but not video. The active parameter count of 4B is comparable to other small-MoE models, but Gemma 4's specific architecture—trained by Google and fine-tuned for instruction following—may give it an edge on multimodal and scientific tasks.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key="$ORCAROUTER_API_KEY",
)
response = client.chat.completions.create(
model="google/gemma-4-26b-a4b-it",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningrepetition_penaltyresponse_formatseedstopstreamstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p| Input / 1M tokens | $0.060 |
| Output / 1M tokens | $0.330 |
| Cache read / 1M | $0.0075 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/google/gemma-4-26b-a4b-itOpen @misc{orcarouter_gemma_4_26b_a4b_it,
title = {Gemma 4 26B A4B API},
author = {Google},
year = {2026},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/google/gemma-4-26b-a4b-it}
}Google. (2026). Gemma 4 26B A4B API. OrcaRouter. https://www.orcarouter.ai/models/google/gemma-4-26b-a4b-it