ModelsQwen3.5-Flash: Hosted 35B-A3B With a Flat-Priced 1M Window
Qwen3.5-Flash is the hosted Qwen3.5-35B-A3B with a 1M default window and one price across all of it. What you are calling, what it costs and when to pick it over 3.6 Flash.
Most cheap models are cheap until the prompt gets long. Providers increasingly price by context tier, so a request that crosses a threshold is billed at a higher rate for every token in it. Qwen3.5-Flash is unusual because, on Alibaba's own price list, it does not do that: one rate covers the whole window up to 1M tokens.
That single detail is the main reason to read past the name. It makes Qwen3.5-Flash one of the more predictable options for long-document work, and it is the clearest difference between it and the newer Qwen3.6-Flash.
What you are actually calling
Qwen's model card for the open Qwen3.5-35B-A3B says it plainly: "Qwen3.5-Flash is the hosted version corresponding to Qwen3.5-35B-A3B". That is a mixture-of-experts model with 35B total parameters and about 3B active per token, using the family's hybrid design of Gated DeltaNet linear attention mixed with full attention layers. It accepts text, images and video.
The hosted version is served on Alibaba Cloud Model Studio with a 1M-token default context. The snapshot currently listed is qwen3.5-flash-2026-02-23, which appears on OpenRouter as qwen/qwen3.5-flash-02-23. A dated snapshot is a frozen build: pin it in production so the model under your prompts does not change without notice. Pinning model versions explains why that matters more than it looks.
Because it is the 35B-A3B underneath, the open model's card is the best guide to capability. Qwen reports SWE-bench Verified 69.2, LiveCodeBench v6 74.6, TAU2-Bench 81.2, OmniDocBench1.5 89.3 and VideoMME with subtitles 86.6 for it. Hosted serving configuration can differ from the reference setup, so treat those as indicative rather than guaranteed for the API.
One price across the whole window
Alibaba lists Qwen3.5-Flash at $0.10 per million input tokens and $0.40 per million output for any request up to 1M input tokens, on its international pricing at the time of writing. Third-party routers set their own rates, so check the endpoint you actually use.
Compare the tiered siblings for a single request with 800K tokens of input and 2K of output:
- Qwen3.5-Flash: about $0.08 input plus well under a cent of output.
- Qwen3.5-Plus, above its 256K tier at $0.50 / $3.00: about $0.40 input.
- Qwen3.6-Flash, above its 256K tier at $1.00 / $4.00: about $0.80 input.
On short prompts the gaps are much smaller. On long ones, Qwen3.5-Flash is a fifth of Plus and a tenth of the newer Flash. If your workload is reading whole contracts, logs or codebases and answering briefly, that ratio decides the bill.
Native 262K, served at 1M
The underlying model's native context is 262,144 tokens. The card says it reaches about 1,010,000 with YaRN scaling. So the 1M hosted window is a stretched one, and a stretched window is not the same as uniform attention across it.
That does not make the long window useless. It means retrieval-style questions over a long document will usually work, while tasks that need reasoning across facts spread through 900K tokens deserve testing before you trust them. Long context degradation describes what to look for. The cheap way to check is to plant known facts at several depths in a real document and ask about each.
Flash or the open 35B-A3B
The weights for Qwen3.5-35B-A3B are Apache 2.0, so self-hosting is a real option. It needs enough memory for all 35B parameters even though only 3B are active, which puts it beyond most single consumer cards at full precision but within reach of one large GPU or a quantised build.
Self-host when data cannot leave your network, when you need a custom fine-tune, or when your volume keeps a GPU busy around the clock. Otherwise, at these prices, the hosted model is very hard to undercut, and you also get the 1M window without configuring YaRN yourself. The Qwen3.5-35B-A3B guide covers the self-hosted side.
Flash 3.5 or Flash 3.6
Qwen3.6-Flash is newer and, below 256K, costs $0.25 / $1.50 on Alibaba's list. That is 2.5 times the input rate and 3.75 times the output rate of 3.5 Flash. Above 256K it costs ten times as much on both.
The 3.6 generation's open models report clear gains in agentic coding over 3.5, so the newer Flash is the natural candidate when the task is code-heavy and prompts stay short. For long-context reading, extraction and summarisation, 3.5 Flash remains the cheaper default and the burden of proof sits with the upgrade. The Qwen3.6-Flash guide covers what is and is not documented about it.
A sensible way to deploy it
- Pin the dated snapshot rather than the moving alias.
- Decide on thinking per route. The family thinks by default; switch it off for extraction and classification and measure the accuracy change.
- Set an explicit output cap. OpenRouter lists a 65,536-token maximum output, which is far more than most answers need and a long way to run if a reasoning trace loops.
- Use it as the first tier in a router, escalating to Plus or a stronger model only when a check fails.
For high-volume reading jobs with short answers, Qwen3.5-Flash is a strong default. For hard reasoning or long agentic coding, start somewhere bigger and use Flash for the steps that do not need it.
Common questions
Which model is Qwen3.5-Flash based on?
Qwen's model card states that Qwen3.5-Flash is the hosted version corresponding to Qwen3.5-35B-A3B, a mixture-of-experts model with 35B total and about 3B active parameters, served with a 1M-token default context.
Does Qwen3.5-Flash charge more for long prompts?
On Alibaba's international price list at the time of writing, no: it has one rate of $0.10 input and $0.40 output per million tokens for requests up to 1M input tokens. Qwen3.5-Plus and Qwen3.6-Flash both step up above 256K.
Should I upgrade from Qwen3.5-Flash to Qwen3.6-Flash?
Only if you measure a quality gain that pays for it. Below 256K the newer model costs 2.5 times as much for input and 3.75 times for output, and above 256K ten times as much, so long-document workloads rarely justify the switch.
