What Is Qwen3-235B-A22B? Cheap Apache MoE (2026)

Qwen3-235B-A22B is an Apache-2.0 mixture-of-experts model from Alibaba's Qwen team that holds 235 billion total parameters but activates only 22 billion per token, and on OpenRouter it prices at roughly $0.09 input and $0.10 output per million. That combination, frontier-class capacity at commodity output pricing, is the reason it keeps coming up in agent-builder conversations, so it is worth pinning down exactly what the model is and what it actually costs to run.

What is Qwen3-235B-A22B (verified 2026-07)

Qwen3-235B-A22B is a mixture-of-experts language model that routes each token through a small subset of its experts, so despite carrying 235 billion total parameters it activates only about 22 billion on any given forward pass. That sparse design is why a model this large can serve at commodity prices.

The "A22B" in the name is the active-parameter count, and the "2507" tag on the current checkpoint marks the July release line. In practice you get most of the reasoning headroom of a 235B dense model while paying inference cost closer to a 22B one. The open weights are the other half of the story.

According to the Qwen3 repository on GitHub, the model is released under the Apache 2.0 license, which places no restriction on commercial use or self-hosting. That matters because most of the cheap Chinese open-weight models buyers ask about carry the same permissive terms, and Apache 2.0 is the most deployment-friendly of them. According to the Apache License, Version 2.0, the grant is permissive and includes an explicit patent license, which is what makes an open-weight model like this safe to ship inside a commercial product rather than only prototype with.

Qwen3-235B-A22B pricing vs other open-weight models

The headline is the output rate. Here is where Qwen3-235B-A22B sits against the other open-weight models we priced on the same routing marketplace, all figures measured via OpenRouter and subject to native re-verification:

The output column is the one that jumps out. At $0.10 per million, Qwen3-235B-A22B undercuts its own sibling coder model by 8x on output and GLM-5 by roughly 19x. It helps to have a first-party yardstick. According to DeepSeek's official API pricing, its V4-Flash endpoint lists at $0.14 input and $0.28 output per million, so the routed Qwen3-235B-A22B output rate lands below even that aggressively-priced first-party number.

One model in this space we could not price directly is Moonshot's Kimi-Dev-72B, an open coding model that was not callable on the routers we tested. We mention it only as sourced context: Moonshot's own materials report it at roughly 60.4% on SWE-bench Verified. We are not presenting that figure as first-hand, and we did not run the model, so treat it as a vendor claim rather than a measured result.

What it cost when we called it (first-hand evidence)

Rate cards are one thing; a billed invoice is another. So we called the model and read the meter. When we called Qwen3-235B-A22B through OpenRouter on 2026-07-10, a short prompt of 36 input tokens returned 66 output tokens in 2.87 seconds and billed $0.00004485. That is the cheapest single run in our open-weight test set, and it is measured, not quoted from a rate card.

To keep the comparison honest we ran the same short prompt across the set the same day. Every number below is measured via OpenRouter routing, and because routing can differ from a model's native first-party endpoint, these carry a needs-native-reverify flag:

Two things surprised us. First, the Qwen run was not only the cheapest but among the fastest, at 2.87 seconds against 6 to 7.6 seconds for the DeepSeek and GLM runs. Second, the billed cost tracked the published rate closely rather than carrying a hidden routing markup, which is not always the case on marketplaces. The served model string came back as qwen3-235b-a22b-07-25, confirming we hit the July checkpoint and not an older build. We flag the routing because a native Alibaba-hosted call could price or latency differently; verify against the first-party endpoint before you budget production on these exact numbers.

Why $0.10 output matters for agent builders

For agent builders the output price is the number that matters, because agents generate far more tokens than they consume. At roughly $0.10 per million output tokens, Qwen3-235B-A22B lets a multi-step agent run its planning and tool-call loops at a fraction of what a frontier closed model would charge for the same volume.

Put concrete numbers on it. A closed frontier model at $10 or $15 per million output makes a chatty agent expensive fast, since every reasoning step, every retry, and every tool-result summary is output you pay for. Drop the output rate to $0.10 and the same loop costs two orders of magnitude less, which changes what is economically sane to build. Long-horizon agents that would be reckless at frontier pricing become routine. That is the practical upshot most pricing tables miss: the cheap number is not the input rate buyers stare at, it is the output rate the agent quietly burns through.

The verdict is narrow and decisive. Choose Qwen3-235B-A22B when your workload is output-heavy and you want open weights you can later self-host, and when Apache 2.0 terms matter for shipping inside a product. Look elsewhere if you need a coding-specialized model, where Qwen3-Coder-Next or a dedicated coder is the better fit, or if your compliance posture requires a native first-party endpoint rather than a routed one until the native rates are verified.

FAQ

What is Qwen3-235B-A22B? It is an open-weight mixture-of-experts language model from Alibaba's Qwen team, released under Apache 2.0. It carries 235 billion total parameters but activates only about 22 billion per token, which is what "235B total / 22B active" means.

How much does Qwen3-235B-A22B cost to run? On OpenRouter it prices at roughly $0.09 per million input tokens and $0.10 per million output tokens. In our live test on 2026-07-10, a 36-input, 66-output call billed $0.00004485 and completed in 2.87 seconds, measured via OpenRouter routing.

Is Qwen3-235B-A22B really open source? Its weights are released under the Apache 2.0 license, which permits commercial use, self-hosting, and modification, and includes an explicit patent grant. That is more permissive than many open-weight releases, and it is why the model is safe to deploy inside a commercial product.

Why is the $0.10 output price notable for a 235B model? Because the mixture-of-experts design activates only 22 billion parameters per token, inference cost sits closer to a 22B model than a 235B dense one. The result is frontier-class capacity at output pricing roughly 100x below a typical closed frontier model.

Are these prices native first-party rates? No. Every price and latency here was measured via OpenRouter routing, not Alibaba's native endpoint, so they carry a needs-native-reverify flag. Confirm against the first-party endpoint before budgeting production spend on these exact figures.

This is part of the open-source AI agents in 2026 hub. For the full rate card across Qwen's lineup, see the Qwen API pricing hub.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing and model benchmarking. Last verified: 2026-07-10.