Loading the catalogue…
Qwen's open-weight multimodal MoE model with 397B total and 17B active parameters, image input and a 262,144-token context.
| Inference Provider | Sovereignty | Runs in | In / out per 1M | Status | Pin |
|---|---|---|---|---|---|
| L3 | €0.60 / 3.60 | Online | |||
| L1 | €0.60 / 3.60 | Online | |||
| L1 | €0.60 / 3.60 | Online | |||
| L0 | €0.41 / 2.72 | Online |
16 use cases classified by the curator. The note says why.
Tool-calling, extended thinking, and a 262K context make it well suited to agent loops that read, edit, and test large codebases, though no coding benchmark is curated yet.
Vision input paired with tool-calling supports the screenshot-observe-act loop that browser agents depend on.
The 262K context window can hold large diffs or multi-file repositories in one pass, and extended thinking supports reasoning over changes.
A 262K context window allows whole long documents to be loaded for question answering without aggressive chunking.
Tool-calling capability combined with extended thinking gives reliable argument selection and call sequencing for function-calling applications.
Native vision capability on a large-scale multimodal base makes it a credible option for visual question answering.
Vision capability supports chart reading and narration; the fit is inferred from general multimodal support rather than chart-specific evidence.
Vision input can read document pages, but no OCR-specific benchmark is curated and dedicated OCR models are typically cheaper per page.
Vision input is supported, but no UI-grounding or screenshot benchmark is curated, so the fit is inferred from general vision capability.
The 262K context handles long inputs well, but at 397B total parameters the cost and latency profile is heavier than smaller summarisation options.
No audio output capability is listed; speech synthesis requires a dedicated TTS model.
No audio input is listed among the capabilities (tool-calling, vision, extended thinking), so the model cannot handle transcription directly.
Sourced numbers, not our own evaluations. Percentiles are against the whole Stav catalogue.
Computed from aa-intelligence-index, gpqa-diamond, aa-engineering-index, aa-coding-index, livecodebench, terminal-bench-v4, ifbench, aa-multilingual-index, aa-lcr.
curl https://api.stav.ai/v1/chat/completions \
-H "Authorization: Bearer $STAV_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-5-397b-a17b",
"reasoning_effort": "high",
"messages": [{"role": "user", "content": "Hello!"}]
}'→ X-Stav-Request-Id · X-Stav-Model · X-Stav-Provider → X-Stav-Sovereignty-Level — the rung the request was served at