Qwen3.8 Omni Flash on QuickSilver Pro
Qwen3.8 Omni Flash is Qwen's first agentic omni model — one endpoint that natively understands image, audio and video alongside text, reasons over what it sees and hears, and drives tools across long workflows. It keeps the family's full 1M-token context window with 131K max output. On QuickSilver Pro it's $0.15 input / $0.47 output per million tokens — a single unified rate across every modality, at Alibaba's own first-party price, reachable on the same OpenAI-compatible key and USD balance as the rest of the catalog.
At a glance
Agentic audio-video understanding on one omni endpoint — image, audio, video and tools on a 1M-token window.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $0.15 | $0.47 | — |
| Alibaba (qwen3.8-omni-flash) | $0.15 | $0.47 | same |
| OpenAI (GPT-4o) | $2.50 | $10.00 | 95% lower |
When to use
Reach for Qwen3.8 Omni Flash when a single request has to reason across modalities: transcribe and summarise a call, pull the key moments out of a long video, caption or QA an image, or run an agent that watches a screen and acts through tools. Every input modality — text, image, audio, video — bills at the same low per-token rate, so mixed-media pipelines stay cheap, and the 1M-token window lets a long recording or a whole document set stay resident instead of being chunked. Alibaba tunes it for agentic delivery, so it plans a task and executes it with tools rather than only describing the content — treat those as vendor claims and match them against your own evals.
When to use something else
For the hardest single-shot text reasoning, step up to Qwen3.8 Max or GLM 5.3. If your workload is pure text with no media, Qwen3.8 Flash Next is cheaper without the omni machinery. Thinking is suppressed by default to keep the flash tier fast; if you need a visible reasoning trace, pass enable_thinking=true and budget the extra output tokens. As a just-released model its behaviour can shift with upstream revisions — pin your evals if you need a frozen target.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-omni-flash",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
Text, images, audio and video, natively in one model — it reasons jointly over what it sees and hears rather than routing each modality to a separate pipeline. Output is text. On QuickSilver Pro every input modality bills at the same $0.15 per million tokens: Alibaba tokenises each modality and charges one unified input price, so there is no separate audio or video rate to budget for.
$0.15 input / $0.47 output per million tokens, with cached input reads at $0.016 — a single unified rate across text, image, audio and video. That matches Alibaba's own first-party list; what QuickSilver Pro adds is one OpenAI-compatible key and one USD balance across the whole catalog, with no separate Alibaba account to provision. Migration is a base_url + key swap.
Yes — the full 1M-token window with up to 131K output tokens per call, and it applies to media as well as text: a long video or audio file, once tokenised, shares the same window. Long-context calls bill the same flat $0.15 per million input tokens.
Yes — Qwen3.8 Omni Flash is an OpenAI-compatible Chat Completions endpoint on QuickSilver Pro (the Responses API works too). Set base_url=https://api.quicksilverpro.io/v1, paste your QSP key, and use model="qwen3.8-omni-flash". Multimodal content parts, streaming, tool calling and usage.cost accounting all work. Thinking is off by default; pass enable_thinking=true to opt into a visible reasoning trace.