Qwen: Qwen3.5-122B-A10B

Qwen3.5-122B-A10B is a text generation model from Qwen, released by the provider on 2026-02-25. It sits in the text category and is described by its provider as a native vision-language model built on a hybrid architecture that combines a linear attention mechanism with a sparse mixture-of-experts design, which the provider states achieves higher inference efficiency. The name reflects that architecture: 122B total parameters with 10B active, so you are calling a large model that only activates part of itself per token.

The model accepts text, image and video as input and returns text only. Its context window is 262,144 tokens, and the maximum output is 65,536 tokens. Because input modalities include images and video alongside text, the same endpoint handles plain prompting, document or screenshot understanding, and video-based questions without switching to a separate vision model. Reasoning output is also part of the picture: the supported parameters include reasoning and include_reasoning, so you can request or surface the model's reasoning content where your application needs it.

Those capabilities point to a fairly wide set of practical uses. Long context plus vision input suits codebase and repository analysis where you paste large amounts of source and ask for explanations or changes, and it suits long document workflows such as contract review, research synthesis or report generation over many pages. Video and image input open up media workflows: describing footage, extracting structured data from frames or screenshots, and answering questions grounded in visual material. Tool calling and structured outputs make it usable as the reasoning core of an agent or a pipeline that must return machine-readable results rather than free text.

You call it through api14's OpenAI-compatible API using the model id "qwen/qwen3.5-122b-a10b". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. In practice that covers sampling control (temperature, top_p, top_k, min_p, the penalty parameters, seed), output control (max_tokens, stop, response_format, structured_outputs), tool use (tools, tool_choice), reasoning control (reasoning, include_reasoning) and inspection of token probabilities (logprobs, top_logprobs, logit_bias).

Two practical constraints are worth keeping in mind when you decide. Output is capped at 65,536 tokens even though the context window is 262,144 tokens, so tasks that need a very long single response may require splitting the work across calls. And the model returns text only, so if your pipeline needs generated images, audio or embeddings, this model is not the right component for that step.