A lossless uncensored version of Qwen3.6 35B A3B that preserves the original model's capabilities while removing refusal behavior. Built for developers, AI researchers, and advanced agent workflows that require unrestricted model outputs without compromising reasoning, coding, multilingual performance, or tool use. The Aggressive variant is fully unlocked and designed to provide direct, complete responses across a wide range of prompts. While the model may occasionally append brief informational disclaimers inherited from the base model's training, these are not refusals and the requested content is still generated in full. Ideal for AI research, security testing, red teaming, agent development, coding assistants, and other advanced AI applications that benefit from maximum output flexibility. 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.
This model is a Mixture-of-Experts variant from the Qwen3.6 series, developed originally by Alibaba Cloud and hosted by provider Obsidian on OrcaRouter. It has 35 billion total parameters, with only…
This model excels at tasks that benefit from a large context window and multimodal understanding. For text, it can summarize or answer questions on documents up to 262,144 tokens long, such as entire books or lengthy reports. For images and videos, it can extract detailed information and combine it with textual context. The uncensored nature allows it to generate creative content without typical safety constraints, useful for story generation, role-play, or other scenarios where restrictive filters are unwanted. Its MoE design provides fast inference relative to dense models of similar total size, making it practical for real-time applications when deployed efficiently. Multilingual capabilities are inherited from the Qwen3.6 family, supporting many languages.
The "uncensored" tag means the model has undergone reduced alignment training compared to standard Qwen3.6 checkpoints. This can result in outputs that are less filtered for harmful, offensive, or controversial content. Users should test the model thoroughly in their specific use case to ensure outputs meet their standards. The lack of censorship can be beneficial for tasks like creative writing, uncensored role-play, or research on sensitive topics where neutral generation is desired. However, it also means the model may produce content that violates typical content policies. OrcaRouter does not claim any additional safety filtering; the model's base behavior is what you get.
If your task involves short inputs (e.g., a few hundred tokens), simple classification, or requires only basic language understanding, smaller and cheaper models like Qwen2.5-7B or GPT-4o-mini may be more cost-effective. This model's strengths are most apparent when handling very long contexts (over 10K tokens) or multimodal inputs. For pure text tasks under 8K tokens with no need for image/video capability, you may save money by using a dedicated small model. Also, if your application requires strict content filtering, the uncensored nature might force you to add an external moderation layer, adding cost.
The model accepts image and video inputs in addition to text. For images, you can provide base64-encoded image data or URLs (via the API's content array with type "image_url"). For videos, the API likely accepts video frames as image sequences or video file URLs; the exact format should be checked in OrcaRouter's documentation. The model processes these visual inputs together with text to generate responses. The context window of 262,144 tokens applies to the combined token count of all input modalities. Note that processing images and videos consumes tokens proportional to the visual resolution and length, so larger inputs will reduce the available text capacity.
No specific benchmark scores were provided for this model by the provider. The Qwen3.6 series generally performs competitively on long-context benchmarks such as L-Eval and RULER, and on multimodal tasks like MMMU and MathVista, but exact figures for this 35B A3B uncensored variant are not published. Users interested in empirical performance should run their own evaluations on representative datasets. The MoE architecture with 3B activated parameters often yields results comparable to dense models of similar active size, while total storage and memory costs are higher due to the full 35B parameters.
Speed and latency depend on several factors: input length, output length, server load, and hardware configuration. Because the model activates only 3B parameters per token, it is expected to be faster than a dense 35B model, with generation rates comparable to models like GPT-3.5 (though no exact figures are provided). OrcaRouter's infrastructure via Obsidian may offer variable latency; users can test with their own workloads. For long-context scenarios (e.g., 100K+ tokens), prefill time increases linearly with input length. Output token generation is typically the main latency contributor. No service-level agreement (SLA) for speed is stated.
The model's strengths include its large 262K context window, which allows processing of entire books or multi-hour video transcripts in one pass. The MoE design provides a good trade-off between capacity and inference speed. Multimodal input enables tasks that combine text and visual data. The uncensored nature allows generation of content that other models might refuse. Multilingual support covers dozens of languages. The pricing is competitive for a model of this capability: $0.31/M input and $4.21/M output, zero markup.
Limitations include the lack of published benchmark numbers, making it harder to gauge relative performance. The uncensored nature may produce undesirable outputs without additional moderation. The 3B activated parameter count means it may not match the raw reasoning power of larger dense models (e.g., GPT-4) on complex logic tasks. Multimodal capabilities may be less refined than dedicated vision-language models. Input modalities like video tokenization could reduce the effective context for text. No streaming or tool-use capabilities are guaranteed. Finally, the dependency on third-party provider Obsidian means availability and uptime are not under OrcaRouter's control.
Pricing is transparent: $0.31 per million input tokens and $4.21 per million output tokens. These are the exact provider rates from Obsidian, with zero markup added by OrcaRouter. Input tokens include all text and visual tokens (for images and videos). Output tokens are the generated text. There is no per-request fee or subscription required. You pay only for the tokens consumed. Pricing is per 1M tokens, so small requests cost fractions of a cent. No free tier or trial is advertised.
No discounts, caching benefits, or volume pricing have been disclosed for this model. The listed rates are flat per token. Unlike some providers that offer reduced prices for cached input tokens, OrcaRouter applies the same input rate regardless of repetition. For very high usage, you may contact OrcaRouter for potential custom arrangements, but no standard discount is documented. The zero-markup policy ensures you are paying exactly what Obsidian charges, with no hidden fees.
Comparable long-context multimodal models from other providers often cost more: for example, OpenAI's GPT-4 Turbo is $10/1M input and $30/1M output, while Claude 3.5 Sonnet is $3/1M input and $15/1M output. This model is notably cheaper at $0.31/$4.21. However, those models offer different capabilities (e.g., tool use, higher reasoning benchmarks). The lower price reflects the MoE architecture and the fact that it is an uncensored third-party hosted model. If you require guaranteed reliability, higher security, or advanced features like function calling, the extra cost may be justified.
To use the model, send a POST request to https://api.orcarouter.ai/v1/chat/completions (or the appropriate endpoint per OrcaRouter's OpenAI-compatible API). Include the header "Authorization: Bearer YOUR_API_KEY". In the request body, set "model": "obsidian/Qwen3.6-35B-A3B". Pass messages in the standard OpenAI format: an array of objects with "role" and "content". For multimodal content, use a content array with type "text" and "image_url" (or video equivalent). Example with cURL: curl https://api.orcarouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $KEY" -d '{"model":"obsidian/Qwen3.6-35B-A3B","messages":[{"role":"user","content":[{"type":"text","text":"Describe this image"},{"type":"image_url","image_url":{"url":"https://example.com/photo.jpg"}}]}]}'.
You can use typical OpenAI-compatible parameters: temperature (default 1.0), top_p (0.9), max_tokens (default varies, recommend setting explicitly), stop sequences, frequency_penalty, presence_penalty, and seed for reproducibility. The model supports streaming via "stream": true to receive token-by-token responses. For multimodal inputs, the API expects content arrays with type "image_url" and optionally "video_url" (check OrcaRouter documentation for exact video format). The context window limit of 262,144 tokens applies to the total number of tokens in messages plus the response. If you exceed the limit, you'll get a context length error; you can adjust "max_tokens" or truncate messages.
To leverage the 262K context, structure your prompts to include the full document or conversation history. Be mindful that every token in the input counts toward cost and limits. For very long inputs, consider chunking or summarizing before sending, although the model is designed to handle the full context. Use the "max_tokens" parameter to control output length. If you encounter timeout errors, enable streaming to get partial results faster. OrcaRouter may have its own timeout settings; contact support if needed. The model does not support system messages in a special way; treat them as a user or assistant message with role "system" (standard OpenAI format).
Migrating from providers like OpenAI or Anthropic involves changing the base URL to https://api.orcarouter.ai/v1 and updating the model ID. If your code uses the OpenAI Python client, set client = OpenAI(api_key="YOUR_KEY", base_url="https://api.orcarouter.ai/v1") and then use client.chat.completions.create(model="obsidian/Qwen3.6-35B-A3B", ...). The message format and parameter names are identical. No changes to prompt logic are required. Note that response object shapes are also compatible, so existing parsing code should work. Test with a small request first to ensure proper authentication and billing.
Compared to Qwen3.6-72B (dense), this model uses MoE and thus has lower inference cost per token but may have slightly lower raw capacity. The 262K context is larger than Qwen2.5's 128K. Compared to Qwen3.6-32B (maybe dense), this model offers multimodal input which those earlier versions may not have. The "uncensored" variant is unique; standard Qwen releases have alignment. If you need strict safety, choose the regular Qwen3.6. In terms of price, this model is cheaper than many dense Qwen models on other platforms.
GPT-4o and Claude 3.5 Sonnet are proprietary models with extensive safety training, higher benchmark scores on reasoning and coding, and support tool use, vision, and large context (200K for Claude). This model offers a larger context (262K) and multimodal input at a significantly lower price. However, it lacks the advanced features, reliability, and performance consistency of those models. For applications where cost is a primary concern and you can accept some quality trade-off, this model is a good alternative. If you need top-tier reasoning or function calling, the premium models are worth the extra cost.
This model is best when you need: (1) a very long context (262K tokens) at low cost; (2) multimodal input (images/video) without paying premium prices; (3) uncensored text generation where content filters would interfere; (4) fast inference relative to total parameter count due to MoE activation. It is a strong contender for research, data extraction, creative writing, and any task that benefits from high context and low per-token cost. If you need advanced features like tool calls, structured outputs, or high reliability, consider other models.
| Input / 1M tokens | $0.310 |
| Output / 1M tokens | $4.21 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/obsidian/Qwen3.6-35B-A3BOpen @misc{orcarouter_qwen3_6_35b_a3b,
title = {Qwen3.6 35B A3B Uncensored (Aggressive) API},
author = {obsidian},
year = {2026},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/obsidian/Qwen3.6-35B-A3B}
}obsidian. (2026). Qwen3.6 35B A3B Uncensored (Aggressive) API. OrcaRouter. https://www.orcarouter.ai/models/obsidian/Qwen3.6-35B-A3B