Qwen3.7 Flash

Overview

Qwen3.7 Flash is the high-throughput tier of the natively vision-language Qwen3.7 series; the current public snapshot is qwen3.7-flash-2026-07-15. In CrossModel, qwen/qwen3.7-flash maps to Bailian's qwen3.7-flash: text, image, and video in, text only out — it does not generate images or video.

Alibaba Cloud's description of this Flash generation centers on two things: multimodal fundamentals and agent execution. Compared with 3.6-Flash it strengthens open-vocabulary recognition, real-world perception, and spatial intelligence; end-to-end execution in multimodal agent settings such as Search Agent and CI Agent is more stable; and multimodal coding was tuned as well. The release shipped without a benchmark table, so everything below is grounded in the specifications and input limits published on the Model Studio model page rather than second-hand scores.

Key capabilities

Qwen3.7 Flash is a hybrid thinking model. enable_thinking defaults to true across the qwen3.7 series, and thinking_budget caps how long the model reasons; Alibaba Cloud recommends streaming output when thinking is on. Fine-tuning is not supported. Current pricing is in the model catalog.

Context and thinking budget

The most concrete change in this generation is the chain-of-thought ceiling: 3.6-Flash allowed 131,072 tokens, 3.7-Flash allows 262,144 — exactly double, and level with the Plus and Max tiers. For multi-step tool calling, UI-operation planning, and long troubleshooting chains, that means the model is no longer forced to cut reasoning short mid-task.

The trade-off is that max input drops from 991,808 to 983,616 tokens once thinking is enabled; the ~8K difference is reserved for the thinking channel. If you intend to fill the 1M window, budget for that gap up front rather than discovering it as a runtime error.

Multimodal input limits

The real dividing line between Flash and Plus is batch size, not fidelity: passed as public URLs or local paths, Qwen3.7 Plus accepts up to 2048 images while Qwen3.7 Flash accepts up to 256; Base64 input is capped at 250 for both. Video limits match — at most 64 videos, each 2 seconds to 2 hours, up to 2GB over a public URL; when a video is supplied as a frame list, the range is 4 to 8000 images.

A few file-level constraints are easy to trip over: for the qwen3.7 series a single image may be up to 20MB by URL, but only 10MB by local path or Base64; the long-to-short side ratio must stay within 200:1; and resolution should stay under 8K (7680×4320), since anything larger raises timeout risk without improving accuracy. Visual tokens and text tokens share one max-input budget, so image count, video length, and text length have to be planned together.

Pricing tiers and cache rules

Qwen3.7 Flash splits long context into three tiers — input ≤32K, 32K–256K, and 256K–1M — priced at roughly 1× / 3× / 6× the base tier, with input and output stepping up together. 3.6-Flash had only two tiers (≤256K and 256K–1M), so a migration cuts both ways: requests under 32K land in a cheaper new tier, while anything past 32K enters the 3× tier earlier than before, which makes context trimming and summarization worth doing for long-running agents.

Cache rules follow Qwen's usual design: an implicit cache hit is billed at 20% of the input rate, explicit cache creation at 125%, and an explicit cache hit at 10%. Batch inference runs at about half the live rate but is offered only in the Beijing region; built-in web search is likewise available only in Beijing and Singapore, not Frankfurt or Tokyo. This article deliberately avoids fixed unit prices — see the model catalog for current rates.

Where it fits in a route

Flash is not a downgraded stand-in for Plus; it is the execution slot on the same route. Give Flash the checkable bulk work — screenshot classification, receipt and document extraction, video summarization, the intermediate steps of search and tool execution — then use JSON schemas, rule checks, or sampled review to isolate low-confidence cases, and escalate only those to Plus for multimodal review or to the text-only Max for final reasoning.

When to use it

• Multimodal agent execution: search, screen reading, tool calls, and intermediate results in Search Agent or CI Agent workflows.
• High-volume visual batches: OCR, screenshot tagging, receipt and document extraction, video summaries — queues whose output can be validated.
• Visual coding: turn screenshots or design mocks into interface code, paired with Function Calling for runnable output.
• Migrating from Qwen3.6 Flash: re-test spatial intelligence, object localization, and end-to-end agent stability, and re-budget context around the new 32K tier.

The CrossModel gateway already serves Qwen3.7 Flash through the OpenAI-compatible /v1/chat/completions interface. Current pricing is in the model catalog.