THE SOURCE IN CONTEXT
Publisher model card with serving links and distinct thinking-mode settings. Use its exact artifact identity alongside the selected engine recipe.
How is the model’s size counted?
The card describes a 125B language-model component with 6B active parameters, plus 51B of n-gram embeddings and a 4B multi-token prediction component. The vLLM recipe uses a 176B summary label. Those counts cover different components; neither label should be substituted for the size of a downloaded artifact. The card lists 48 layers and a repeated hybrid of Gated DeltaNet and sparse attention.
Source: Qwen · pinned document
Is the open checkpoint natively a million-token model?
The published native window is 262,144 tokens. Reaching 1,000,000 tokens requires the documented extension configuration; the hosted Qwen3.8-Flash product and the open Flash-Next checkpoint also have different production features. Record the actual RoPE settings and total input-plus-output budget before evaluating long tasks.
Source: Qwen · pinned document
How do thinking and non-thinking settings differ?
Thinking mode uses temperature 1.0, top_p 0.95 and presence_penalty 0.0; non-thinking uses 0.7, 0.80 and 1.5 respectively. Both recommendations retain top_k 20, min_p 0.0 and repetition_penalty 1.0. The card also preserves historical thinking blocks by default. Changing that conversation policy changes both the prompt and cache behavior.
Source: Qwen · pinned document
How should its benchmark table be read?
The coding evaluations specify harnesses, sampling and context. In particular, DeepSWE reports the better result from Claude Code and mini-SWE-agent rather than one universally shared harness. For vision tables, pass@3 and the mean over three attempts are different statistics. Keep the footnote for the selected task instead of treating every published cell as directly comparable.
Source: Qwen · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| LM component / active / n-gram / MTP | 125B / 6B / 51B / 4B |
|---|---|
| Layers | 48 |
| Native / extended context | 262144 / up to 1000000 tokens |
| Thinking sampling | temperature 1.0; top_p 0.95; top_k 20 |
| Non-thinking sampling | temperature 0.7; top_p 0.80; presence_penalty 1.5 |
AI-assisted source reading and original editorial synthesis; no independent model execution.