A lossless uncensored version of Gemma 4 26B A4B designed to preserve the original model's capabilities while minimizing refusal behavior. Optimized for developers, AI researchers, and advanced applications that require high-quality responses with minimal restrictions. The Balanced variant is recommended for most workloads, delivering complete answers while maintaining stable reasoning and natural conversational behavior. In some sensitive scenarios, the model may briefly frame its reasoning before providing the full response, but it is designed to avoid withholding content. Compared to more aggressive uncensored variants, Balanced offers more consistent sampling, stronger long-context stability, and reduced topic drift across extended conversations. Well suited for creative writing, roleplay, multilingual assistants, long-context reasoning, and general-purpose AI applications where quality, coherence, and reliability are the primary priorities. Access to this model is gated and intended for security researchers, red teams, AI safety researchers, and other qualified professionals conducting legitimate research, evaluation, and testing.
Gemma 4 26B A4B Uncensored is a mixture-of-experts (MoE) language model from Google’s Gemma 4 series, hosted by provider Obsidian and accessible through OrcaRouter. It has 26 billion total parameters…
Gemma 4 26B A4B Uncensored excels in tasks that require long-range reasoning over large contexts, such as book summarization, multi-turn conversations with extensive history, and codebase analysis. Its MoE architecture makes it efficient for batch processing where many prompts are handled concurrently. Multimodal capabilities let it reason about images—for example, interpreting graphs, memes, or product photos. The uncensored behavior is ideal for creative writing that explores mature themes, adult roleplay, or generating fiction without content restrictions. It also performs well on standard NLP benchmarks for reasoning, math, and coding, similar to other Gemma 4 variants.
If your task does not require long context (e.g., under 8K tokens) or multimodal input, you may be better served by a smaller dense model such as Gemma 2 9B or Llama 3 8B, which are faster and cheaper per token. For tasks that need safety filters, the uncensored model could produce inappropriate outputs—choose a safety-tuned model instead. Also, if you need extremely low latency for real-time chat, this MoE model may be slower than a small dense model because of the expert routing overhead. Consider your throughput needs: for high-volume, short-context requests, a cheaper model like Gemma 2 27B (dense) might be more cost-effective.
The mixture-of-experts design activates only a subset of parameters per token (4 billion out of 26 billion total). This reduces the computational cost per forward pass compared to a dense 26B model, allowing higher throughput on the same hardware. However, the routing introduces slight latency overhead per token and can lead to expert load imbalance if prompts are highly specialized. In practice, MoE models like this one offer a good trade-off: competitive quality for a fraction of the FLOPs. The 4B active parameters are comparable to a 4B dense model in terms of per-token compute, yet the model benefits from the knowledge stored across 26B total parameters.
Yes, the model accepts both text and image inputs. You can send a single image or multiple images in a request, alongside text instructions. The images are encoded and processed together with the text within the 262,144-token context window. This enables visual question answering, document understanding, and tasks that require analyzing charts, diagrams, or photographs. The model uses a vision encoder (typically a ViT-like component) to convert images into tokens. While the exact details of the architecture are not publicly documented, it supports standard image formats. Note that images consume tokens proportional to their resolution, so high-resolution images will reduce the available text context.
As a Gemma 4 variant, the base model (with safety tuning) has shown strong performance on reasoning benchmarks like MMLU, GSM8K, and coding tasks such as HumanEval and MBPP. The uncensored version likely retains these capabilities since the core model weights are unchanged. However, specific benchmark scores for this uncensored variant have not been released. Users can expect competitive results in arithmetic reasoning, code generation, and multilingual tasks. The 262K context window also allows superior performance on long-document retrieval and summarization compared to models with shorter context. For multimodal benchmarks, the model should handle visual question answering datasets adequately.
Latency for the Gemma 4 26B A4B model is generally lower than that of a dense 26B parameter model because only 4B parameters are active per token. However, the MoE routing mechanism adds a small overhead to each generated token, so the total latency per token may be slightly higher than a pure 4B dense model. For batch inference with long sequences, the throughput can be higher due to reduced memory bandwidth demands. On OrcaRouter, latency will also depend on the underlying hardware provisioned by the Obsidian provider. Real-world performance should be tested with representative workloads to determine if it meets your speed requirements.
Despite its strengths, this model has limitations. The 4B active parameters mean that for very complex reasoning or domain-specific knowledge, it may underperform larger models like Gemma 4 47B (dense) or frontier models. The uncensored nature means outputs may be toxic, biased, or misaligned with human values if used without moderation. It may also generate harmful content that violates OpenAI's usage policies when accessed through OrcaRouter—users are responsible for compliance. Additionally, the MoE architecture can lead to inconsistent quality across tokens if expert routing is suboptimal. Finally, multimodal understanding, while present, may not match dedicated vision-language models.
Pricing for this model is determined by the provider Obsidian and passed through by OrcaRouter with zero markup. You pay $0.25 per million input tokens and $2.90 per million output tokens. Input tokens include both text and image tokens (images are tokenized before processing). Output tokens cover the generated response. There are no additional fees for API calls or context window usage beyond the per-token cost. This pricing is competitive for a 26B MoE model, especially given the large context window. Charges are calculated per request and appear on your OrcaRouter billing statement.
Output tokens are significantly more expensive than input tokens ($2.90 vs $0.25 per million). This is typical for language models because generating tokens requires more computation than processing input. To minimize costs, you can structure prompts to reduce the number of output tokens—for example, by asking for shorter responses or using system instructions to limit verbosity. Additionally, since input costs are low, you can afford to include large context windows (up to 262K tokens) without breaking the bank. If your use case involves many short prompts but long responses, output token cost will dominate.
OrcaRouter does not provide built-in caching for this model. The pricing is per-token and per-request. However, you can implement client-side caching of common input sequences to avoid re-sending identical prompts. Additionally, if you repeat the same system message across many conversations, you might consider including it in a long context and reusing the same session via the chat completions API to avoid redundant input tokens. Some providers may offer prompt caching on their own, but Obsidian's pricing as listed does not include a caching option. Check the provider's documentation for any potential discounts on repeated prompts.
To use this model, send requests to the OpenAI-compatible endpoint https://api.orcarouter.ai/v1. Set the model parameter to "obsidian/gemma-4-26B-A4B". Your API key from OrcaRouter should be included in the Authorization header as Bearer token. The API supports both text completions and chat completions endpoints. For chat, use the /v1/chat/completions endpoint with messages array including user, assistant, and system roles. For multimodal requests, include image URLs or base64 data in the content field. An example request body in Python would use the openai library with base_url set to OrcaRouter's endpoint and the model ID as above.
The API supports standard OpenAI parameters: temperature (0-2, default 1), top_p (0-1), max_tokens (up to context window), presence_penalty, frequency_penalty, stop sequences, and n (number of completions). For multimodal, the content field accepts an array with type "text" and type "image_url". You can also set stream=true for streaming responses. The model supports system messages. Context window is 262,144 tokens including input and output. Not all parameters may be supported exactly as documented; check the OrcaRouter documentation for any provider-specific limitations. For best results, set temperature between 0.7 and 1.0 for creative tasks, and lower for deterministic outputs.
Migration is straightforward due to the OpenAI-compatible API. If you are currently using the OpenAI Python client, change the base_url to https://api.orcarouter.ai/v1 and set your API key to your OrcaRouter key. Update the model parameter from the previous model name to "obsidian/gemma-4-26B-A4B". No other code changes are required for most use cases. For multimodal requests, the image format may differ; use the same structure as OpenAI's vision API. Test a few requests to ensure compatibility. Note that rate limits and error handling may differ; consult OrcaRouter's documentation for best practices.
The base URL for all API calls is https://api.orcarouter.ai/v1. The exact model identifier to use in requests is "obsidian/gemma-4-26B-A4B". This ID must be passed as the model parameter in chat completions or completions calls. There is no alternative model ID for this variant. Ensure you include the full prefix 'obsidian/' to route correctly to the Obsidian provider. If you attempt to use a generic 'gemma-4-26B-A4B' ID without the provider prefix, it may not resolve. The provider prefix is necessary because OrcaRouter supports multiple providers offering the same base model.
Within the Gemma 4 family, Google offers both dense and MoE variants. The dense version (e.g., Gemma 4 47B) has all parameters active, providing higher quality per token but at higher computational cost. The MoE version with 26B total and 4B active offers a sweet spot for cost-efficiency. Compared to Gemma 4 2B or 9B dense, this model has broader knowledge due to larger total parameter count. Among MoE models, Gemma 4 26B A4B is smaller than larger MoE models like Mixtral 8x22B (141B total, 39B active). The uncensored aspect is unique to this Obsidian variant; other Gemma 4 models include safety tuning.
Uncensored models like those from the Llama 3-Uncensored or Wizard series typically remove safety filters from a base model. This Gemma 4 variant is one of the few MoE uncensored options, offering a larger total parameter count (26B) with lower active parameters (4B). Compared to Llama 3 70B Uncensored, it is much smaller and cheaper but may have lower ceiling on complex tasks. Its 262K context window is significantly larger than most uncensored models, which often cap at 8K-32K. The multimodal support is also a differentiator: most uncensored models are text-only.
GPT-4o and Claude 3.5 Sonnet are larger, proprietary models with extensive safety tuning and multimodal capabilities. They generally outperform Gemma 4 26B A4B on benchmarks and real-world tasks, especially in reasoning, creativity, and consistency. However, they are much more expensive per token (GPT-4o: $5/M input, $15/M output; Claude: $3/M input, $15/M output). The Gemma model is ideal for cost-sensitive applications needing large context but without safety restrictions. It will not match the frontier models on subtle instruction-following or factual accuracy, but may be sufficient for many generative tasks.
Choose this model over a larger model (like GPT-4o or Claud Opus) when: (1) you need a very large context window (262K) without paying premium prices; (2) you require uncensored outputs for allowed use cases; (3) your workload is high-throughput and the MoE cost efficiency matters; (4) your tasks are within the capability of a 4B active parameter model. Avoid this model if you need the highest quality for critical decisions, strict safety alignment, or extremely complex reasoning. For many common tasks like summarization, translation, or Q&A on long documents, this model provides a strong price-performance ratio.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key=os.environ["ORCAROUTER_API_KEY"],
)
response = client.chat.completions.create(
model="obsidian/gemma-4-26B-A4B",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)include_reasoningreasoningmax_tokenstemperaturetop_ptop_kmin_ppresence_penaltyfrequency_penaltystopseedresponse_formatstructured_outputstoolstool_choice| Input / 1M tokens | $0.250 |
| Output / 1M tokens | $2.90 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
What developers are saying this week
GET /api/public/models/obsidian/gemma-4-26B-A4BOpen @misc{orcarouter_gemma_4_26b_a4b,
title = {Gemma4 26B A4B Uncensored (Balanced) API},
author = {obsidian},
year = {2026},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/obsidian/gemma-4-26B-A4B}
}obsidian. (2026). Gemma4 26B A4B Uncensored (Balanced) API. OrcaRouter. https://www.orcarouter.ai/models/obsidian/gemma-4-26B-A4B