Qwen3.8 2.4T A95B

This model was retired on 6 October 2026

Requests for qwen/qwen3.8-2.4t-a95b now return a 410 model_retired error; your request is not silently routed to another model.

Use qwen/qwen3.8-max instead: the same Qwen3.8 generation at the same input and output rates, with a 1,000,000-token context window and every feature on this page, plus image and video input. Moving over has two differences: cached input costs $0.25 per 1M instead of $0.20, and output is capped at 131,072 tokens. When you change the model identifier, lower max_tokens too if yours is above that.

The rest of this page describes how the model behaved while it was available.

Qwen3.8 2.4T A95B was the open-weight release of the Qwen3.8 generation: a sparse Mixture-of-Experts model that ran 95 billion of its 2.4 trillion parameters per token. It was built for coding, research and long-horizon agent workflows.

Model	Context	Input / Output ($/1M)	Cached input ($/1M)
qwen/qwen3.8-2.4t-a95b	262,144	2.00 / 6.00	0.20

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.8-2.4t-a95b",
    "messages": [
      { "role": "user", "content": "Find the race condition in this function." }
    ],
    "max_tokens": 8192
  }'

Successor: Qwen3.8-Max

Max is the model that replaces the open-weight release:

	qwen/qwen3.8-2.4t-a95b	qwen/qwen3.8-max
Release	Open weight	Closed, provider-hosted
Context	262,144	1,000,000
Max output	Bounded by the context window	131,072
Input types	Text only	Text, image, video
Can reasoning be turned off	No	Yes
Input / Output ($/1M)	2.00 / 6.00	2.00 / 6.00

To move over, change the model id to qwen/qwen3.8-max and keep max_tokens at or below 131,072; the old id is not forwarded to its successor.

Reasoning is always on

The model reasons step by step before answering and returns that chain separately in reasoning_content. content carries the final answer.

This cannot be turned off. Requests that try to disable reasoning are rejected by the provider. The :fast suffix, "reasoning": false and similar controls that work on other models do not silence reasoning here.

The practical consequence: reasoning tokens are output and are billed inside completion_tokens. Even a short question can spend a few hundred reasoning tokens, so keep max_tokens generous. If it is too low the budget is consumed during reasoning and you are left with an empty content and finish_reason: "length".

reasoning_effort is not supported on this model

The provider appears to accept seven levels, but the levels have no measurable effect: on the same prompt, xhigh can produce shorter reasoning than none, and repeating a single level varies more than the difference between levels.

Rather than offer a control that does nothing, LLMTR rejects the parameter with a 400. For models that genuinely implement levelled control, see Reasoning Effort.

Tool calling

Tool calling is native. All three forms of tool_choice work — auto, required and a named function — and the returned tool_calls arguments are valid JSON.

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.8-2.4t-a95b",
    "messages": [{ "role": "user", "content": "Look up order 12345." }],
    "tools": [{
      "type": "function",
      "function": {
        "name": "lookup_order",
        "description": "Look up an order by id",
        "parameters": {
          "type": "object",
          "properties": { "orderId": { "type": "string" } },
          "required": ["orderId"]
        }
      }
    }],
    "tool_choice": "auto"
  }'

You can feed the tool result back with a role: "tool" message and continue the conversation. See Tool Calling for details.

Structured output

response_format accepts both json_object and json_schema:

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "capital",
      "schema": {
        "type": "object",
        "properties": {
          "country": { "type": "string" },
          "capital": { "type": "string" }
        },
        "required": ["country", "capital"]
      }
    }
  }
}

Prompt caching

Repeated prompt prefixes are cached automatically; no extra parameter is needed. The cache warms on the first call and hits on subsequent ones.

Tokens read from cache are reported in usage.prompt_tokens_details.cached_tokens and billed at $0.20 per 1M instead of the input rate — one tenth of the normal input price.

Measured example: an identical 19,627-token prefix returned 17,952 tokens from cache on the second call. If you work with a long, fixed system prompt, keeping that prefix byte-identical across requests lowers cost noticeably.

Limits

• Context is 262,144 tokens. A request whose input plus output exceeds it is rejected with a 400.
• There is no separate output ceiling. No limit is applied to max_tokens beyond the context window.
• Text only. Image, audio and video input are not accepted; requests carrying an image are rejected.

Data residency

This model runs on third-party infrastructure outside Turkey. Prompts and completions may be logged by the provider. It is not an LLMTR model hosted in Turkey. For Turkey-hosted models see LLMTR Qwen.