To reproduce Qwen3.5-Plus's million-token context and Agent workflows, fix the model snapshot, deployment region, input modalities, and tool capabilities together.
Suitable tasks: long-document/codebase analysis, image or video understanding, structured extraction, function calling, and web-search workflows.
Unsuitable tasks: fine-tuning the model, directly transferring web-search capabilities from one region to another, or requiring audio input not listed on the official page.
Applicable model versions: qwen3.5-plus defaults to qwen3.5-plus-2026-02-15; the page also lists the qwen3.5-plus-2026-04-20 snapshot.
Applicable client, Agent, or API: Alibaba Cloud Model Studio; compatibility interfaces and specific tool orchestration are subject to the service-side documentation.
Recommended inference tier and parameters: the official page does not disclose recommended temperature/top_p values; start with the service defaults, fix the snapshot, and set the output limit to no more than 65,536 tokens. For stable reproduction, do not mix the unfixed qwen3.5-plus alias with a fixed snapshot.
The following configuration template is organized from the official capability table; it is not the official source code. Replace region, snapshot, and the tool allowlist for the actual tenant.
Model: qwen3.5-plus-2026-02-15
Region: <actual Model Studio deployment region>
Input: text + image/video (choose according to the task)
Output: text
Context window: no more than 1,000,000 tokens
Maximum input: no more than 991,808 tokens
Maximum output: no more than 65,536 tokens
Tool allowlist:
1. function calling: register only the functions required to complete the task, and specify the type, range, and failure return for each parameter.
2. structured outputs: specify a fixed JSON schema; return null for fields that cannot be confirmed, and do not guess.
3. web search: first confirm that the capability table for the current deployment region is marked “supported,” then require the model to cite sources for external facts.
System constraints:
- First determine the input modality and task objective, then select tools.
- Before each tool call, state its purpose, inputs, and expected result; when a tool fails, return a diagnosable error and stop retrying.
- The final output must contain only the agreed schema, along with the snapshot, region, and tool execution status.Create or select a fixed qwen3.5-plus-2026-02-15 endpoint in Model Studio, and record the region and request time.
Using the same text, image, or video input, run a no-tool baseline once and record input/output token counts and latency.
Enable only one function, repeat the task, and check whether the function arguments conform to the schema; then add structured output and web search step by step.
For inputs exceeding 128k and 256k, record the billing tier and cache-hit status separately; do not conflate capability tables from different regions.
If switching to qwen3.5-plus-2026-04-20, rerun the complete baseline and compare Agentic coding, speed, and long-context performance separately.
The official capability table lists text, image, and video input and text output; function calling, structured outputs, and prefix continuation are all supported.
The China (Beijing) region table lists web search, context caching, and batch inference as supported; web search, caching, and batch capabilities in the international/global tables vary with the deployment scope.
The context window is 1,000,000, the maximum input length is 991,808, and the maximum output length is 65,536; in thinking mode, the maximum input is 983,616 and the maximum chain-of-thought length is 81,920.
In China (Beijing), the list price is tiered by input length: for no more than 128k, input is RMB 0.8 per million tokens and output is RMB 4.8 per million tokens; for 128k–256k, it is 2/12; for 256k–1m, it is 4/24. The page also lists international and global prices; the endpoint region determines the actual price.
The official description of the 2026-04-20 snapshot says that it improves Agentic coding and inference speed over 2026-02-15 and is suitable for coding agents, production workflows, and high-throughput scenarios; this is a version note, not an independent controlled evaluation.
“Web search supported” is listed by region/deployment scope; support in the China region should not be taken to imply that international or global endpoints necessarily support it.
The page does not provide a complete system prompt, tool schema, or recommended temperature. The prompts and parameter limits in this article are a reusable organization of the material and should not be presented as official source text.
The official page does not list audio input; for audio, choose a model or conversion workflow that explicitly supports it.
2026-04-20 is a separate snapshot; combining it with results from the default alias would undermine reproducibility.
The official documentation describes the default version as “functionally equivalent to the snapshot model qwen3.5-plus-2026-02-15” and lists the 2026-04-20 snapshot separately; the configuration above follows that boundary.
Qwen3.5 Plus