Qwen 3.5 Flash
Qwen 3.5 Flash is Alibaba's production-hosted multimodal model built on a hybrid linear-attention MoE architecture, offering a context window of 1M tokens and sub-second responsiveness for high-throughput agentic workloads.
import { streamText } from 'ai'
const result = streamText({ model: 'alibaba/qwen3.5-flash', prompt: 'Why is the sky blue?'})Playground
Try out Qwen 3.5 Flash by Alibaba. Usage is billed to your team at API rates. Free users (those who haven't made a payment) get $5 of credits every 30 days.
Providers
Route requests across multiple providers. Copy a provider slug to set your preference. Visit the docs for more info. Using a provider means you agree to their terms, listed under Legal.
P50 throughput on live AI Gateway traffic, in tokens per second (TPS). Visit the docs for more info.
P50 time to first token (TTFT) on live AI Gateway traffic, in milliseconds. View the docs for more info.
Direct request success rate on AI Gateway and per-provider. Visit the docs for more info.
More models by Alibaba
| Model |
|---|
| 240K | 3.2s | 50tps | $1.30/M | $7.80/M | Read: $0.26/M Write: $1.63/M | — | — |  | 04/20/2026 | ||||
| 1M | 0.7s | 55tps | $0.50/M | $3.00/M | Read: $0.1/M Write: $0.63/M | — | — |  | 04/02/2026 | ||||
| 1M | 1.6s | 110tps | $0.40/M | $2.40/M | Read: $0.04/M Write: $0.5/M | — | — | | 02/16/2026 | ||||
| 256K | 1.0s | 62tps | $0.50/M | $1.20/M | — | — |   | 07/22/2025 | |||||
| 262K | 0.1s | 86tps | $0.07/M | $0.46/M | Read:$0.6/M Write:— | — | — |    +1 | 04/01/2025 | ||||
| 131K | 0.2s | 282tps | $0.10/M | $0.30/M | Read:$0.14/M Write:— | — | — | +1 | 04/01/2025 |
About Qwen 3.5 Flash
Qwen 3.5 Flash is built on Alibaba's fifth-generation Qwen3.5 architecture, which combines Gated DeltaNet linear attention with sparse mixture-of-experts layers in a 3:1 linear-to-full attention ratio. This design allows the model to process very long documents and codebases efficiently while keeping inference costs low, the hosted Flash tier makes contexts of 1M tokens the default rather than an opt-in premium.
The model handles text, images, and video natively in a single forward pass, without requiring separate vision adapters. That native multimodality makes it well-suited for workflows that mix screenshot analysis, document review, and code generation in the same conversation. Structured outputs, tool calling, and seed-based reproducibility are all supported out of the box.
Qwen 3.5 Flash ships with configurable reasoning depth, letting callers dial up or down the amount of internal chain-of-thought the model performs before responding. At lower reasoning settings the model behaves like a fast instruction-follower; at higher settings it performs multi-step decomposition suitable for mathematical problem solving or complex agentic tasks.
What To Consider When Choosing a Provider
- Configuration: For latency-sensitive pipelines, compare time-to-first-token across available providers using the AI Gateway playground before committing to a routing configuration.
- Zero Data Retention: AI Gateway does not currently support Zero Data Retention for this model. See the documentation for models that support ZDR.
- Authentication: AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.
When to Use Qwen 3.5 Flash
Best For
- Whole-codebase and long-PDF processing: Handling entire repositories or long reports in a single request using the default context of 1M tokens
- Fast agentic tool loops: Low-cost structured JSON responses for agents that chain many tool calls
- Multimodal conversation threads: Pipelines where text, screenshots, and short video clips arrive in the same thread
- Latency-sensitive reasoning: Applications that need reasoning capability but can't tolerate the cost of the Plus tier
Consider Alternatives When
- Maximum reasoning depth: Consider Qwen3.5 Plus for heavier analytical workloads when cost is secondary
- Lowest text-only pricing: A dedicated text model is cheaper for pipelines that never need vision
- Image or video generation: This model understands multimodal inputs but doesn't generate images or video
Conclusion
Qwen 3.5 Flash delivers Alibaba's fifth-generation multimodal reasoning at a cost point suited for production scale, with a context of 1M tokens that eliminates most RAG pipeline overhead. For teams building document-heavy or agentic applications on Vercel, it occupies the efficiency end of the Qwen3.5 lineup without sacrificing vision support.
Frequently Asked Questions
What architecture powers Qwen 3.5 Flash?
It uses a Gated DeltaNet plus sparse mixture-of-experts design with a 3:1 linear-to-full attention ratio, enabling efficient processing of very long sequences at lower compute cost than dense transformer models.
Can Qwen 3.5 Flash analyze video clips?
Yes. The model natively accepts video inputs alongside text and images, allowing you to include short video segments in the same prompt as text instructions without preprocessing.
How does the context of 1M tokens affect RAG architecture decisions?
For many document retrieval tasks the full context window eliminates the need for a separate vector search layer, since entire documents or codebases can be passed directly. However, chunking and retrieval still benefit latency and cost for very large corpora.
Does Qwen 3.5 Flash support tool calling?
Yes. Tool calling, structured JSON outputs, and function-calling patterns are fully supported across all AI Gateway interfaces.
What does the configurable reasoning parameter do?
Callers can adjust how much internal chain-of-thought computation the model performs before responding. Lower settings optimize for speed; higher settings improve accuracy on multi-step reasoning tasks at the cost of added latency.
What is the difference between Qwen 3.5 Flash and Qwen3.5-Plus?
Flash is the cost-optimized, lower-latency variant built on the 35B-A3B architecture, while Plus is the higher-capability tier suited for more demanding reasoning and visual analysis tasks. Both share the context window of 1M tokens.
Is Qwen 3.5 Flash suitable for production agentic workflows?
Yes. The model was specifically designed for agentic use: it supports adaptive tool use, structured outputs, and the long context required to maintain agent state across many tool-call turns.