Qwen: Qwen3.8 2.4T A95B

Qwen3.8 2.4T A95B is an open-weight sparse mixture-of-experts model from Qwen, and it is the open-weight variant of Qwen3.8 Max. The name reflects its architecture: 95 billion active parameters out of 2.4 trillion total, so only a fraction of the network is engaged per token while the full model still holds a very large pool of weights. It sits in the text category, takes text in, and returns text out.

The headline number for most integrations is the context window of 1,048,576 tokens, which is large enough to hold entire codebases, long document sets or extended conversation histories in a single request. Output is capped at 131,072 tokens, so long generations are possible alongside long inputs. The model supports reasoning behaviour, including a reasoning parameter, a reasoning_effort control and an include_reasoning flag, which lets you decide whether reasoning content is surfaced. It also supports structured outputs and response_format, so you can constrain replies to a schema rather than parsing free text.

Those capabilities point to a fairly specific set of uses. Long-context analysis is the obvious one: feeding whole repositories, contract bundles, log archives or research corpora and asking questions across all of it without a retrieval layer. Because tools and tool_choice are supported, the model can act as the reasoning core of an agent that calls functions or external APIs. Structured outputs make it a reasonable fit for extraction pipelines, where you need consistent JSON from messy input. The reasoning controls suit tasks where you want the model to work through a problem step by step, and the sampling parameters give you room to tune behaviour for either deterministic or more exploratory generation.

Calling it through api14 works the same way as any other model on the platform, since the API is OpenAI-compatible. You send your request to the chat completions endpoint with the model id "qwen/qwen3.8-2.4t-a95b" and your messages in the usual format. The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. If you already have code written against the OpenAI SDK, pointing the base URL at api14 and swapping the model id is generally the whole change.

Two practical notes. First, this is a text-only model: there is no image or audio input, so anything multimodal needs a different model. Second, the mixture-of-experts design means the active parameter count is much smaller than the total, but a 2.4 trillion parameter model is still a large one, and the 1M token context means requests can carry a lot of input. For short, simple tasks where latency or throughput matter more than depth, a smaller model will usually be the better choice.