Loading page
Alibaba
Qwen3 235B A22B 2507
Qwen3 235B A22B 2507 is an updated release in Alibaba's Qwen3 model family, using a mixture-of-experts design with roughly 235B total parameters and about 22B active per token.
From
$0.020 / megapixel
across 4 providers
API Pricing
| Provider | Price /MP | Alt /MP |
|---|---|---|
| $0.020/MP | - | |
| $0.087 | $0.350 | |
| $0.200 | $0.600 | |
| $0.200 | $0.800 |
Prices updated daily. Last check: Sep 21, 2026
Qwen3 235B A22B 2507 pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
12.0 / 100
Math
71.7 / 100
Reasoning & Knowledge
- MMLU-Pro82.8%
- GPQA Diamond75.3%
- Humanity's Last Exam11.1%
Coding
- LiveCodeBench52.4%
Math
- AIME 202571.7%
- AIME71.7%
- MATH-50098.0%
Agentic & Tool Use
- Terminal-Bench Hard15.2%
- τ²-bench33.3%
Instruction & Long Context
- IFBench46.1%
- Long-Context Reasoning33.9%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Alibaba
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Mixture-of-experts design activates roughly 22B of about 235B parameters per token, keeping per-token compute lower than a dense model of comparable total size
- A dated 2507 refresh of the earlier Qwen3 235B A22B checkpoint, so improvements arrive without a change in model naming or integration path
- Available from multiple third-party inference providers, allowing price and latency comparison rather than dependence on one endpoint
- Part of the broader Qwen3 family, making it straightforward to move between larger and smaller sizes as workload requirements change
- Qwen3 models are commonly used for multilingual workloads, including Chinese-English tasks
- Large total parameter budget gives it more capacity than the mid-sized members of the Qwen3 lineup
Limitations
- We do not have measured output throughput or time-to-first-token data for this model, so speed must be evaluated per provider
- Total parameter count means substantial GPU memory is required for self-hosting, even though only a subset of experts is active per token
- Context window, modality support, and tool-calling behavior are not recorded in our metadata and vary by serving provider
- The 2507 dated suffix can be confused with the earlier Qwen3 235B A22B release when pinning a specific version in production
- Multiple providers serving the same weights means output behavior can differ with quantization and sampling defaults
Key Features
•Mixture-of-experts architecture with approximately 22B active parameters per token
•Roughly 235B total parameters across expert layers
•2507-dated refresh of the Qwen3 235B A22B checkpoint
•Member of Alibaba's Qwen3 model generation
•Multilingual text generation, including Chinese and English
•Multi-provider availability through third-party inference APIs
•Text-in, text-out chat completion interface as served by hosted providers
About Qwen3 235B A22B 2507
Qwen3 235B A22B 2507 is a large language model from Alibaba's Qwen team, part of the Qwen3 generation. The naming convention describes its architecture: roughly 235 billion total parameters with approximately 22 billion active per forward pass (the "A22B" designation), indicating a sparse mixture-of-experts (MoE) layout rather than a dense model. The "2507" suffix marks it as a dated refresh of the original Qwen3 235B A22B release, following Alibaba's practice of shipping incremental updates to an existing checkpoint name instead of renaming the model.
The MoE structure is the defining technical characteristic here: only a fraction of the total parameter count is engaged for any given token, so serving cost and latency track the active parameter count more closely than the headline size. That makes it a large-capacity model that is comparatively cheaper to run per token than a dense model of similar total size, though it still requires enough memory to hold all experts. Qwen3 models are generally positioned as multilingual, with coding, math, and instruction-following among the commonly cited focus areas for the family.
In practice, Qwen3 235B A22B 2507 is served through multiple inference providers rather than a single first-party API, which means throughput, latency, maximum context length, and supported features can differ meaningfully between endpoints. Buyers comparing it against other large open-weight models typically weigh it on cost per token at a given throughput level, and against Alibaba's own smaller Qwen3 variants when the workload does not need the larger model's capacity. Our benchmark record for this model currently has no measured throughput or time-to-first-token figures, so consult per-provider numbers where available.
Common Use Cases
Qwen3 235B A22B 2507 suits workloads that need a high-capacity model but are sensitive to per-token serving cost, since the mixture-of-experts design engages only about 22B parameters per token. Typical fits include general assistant and chat backends, code generation and code explanation, longer-form drafting and summarization, and multilingual tasks where Chinese-English quality matters. It is also a common choice for teams that want a large open-weight alternative to proprietary APIs, either through a hosted provider or on their own GPU fleet, and for evaluation pipelines that compare several large models side by side. For high-volume classification, extraction, or routing work where the marginal quality of a 235B-class model is not needed, a smaller Qwen3 variant is usually the more economical option.
Frequently Asked Questions
How much does Qwen3 235B A22B 2507 cost to use?
Pricing depends on which provider you use and how you buy capacity — per-token serverless inference, dedicated endpoints, or self-hosting on rented GPUs all price differently, and rates change frequently. Check the pricing table on this page for current per-provider figures.
What is Qwen3 235B A22B 2507 best used for?
It is generally used for general-purpose chat and assistant backends, coding help, longer-form writing and summarization, and multilingual tasks. Its mixture-of-experts design makes it attractive when you want large-model capacity without paying dense-model compute for every token.
What do the numbers in the name mean?
235B refers to the approximate total parameter count and A22B to the roughly 22 billion parameters active per token, which indicates a sparse mixture-of-experts architecture. The 2507 suffix identifies this as a dated update to the original Qwen3 235B A22B release.
How does it compare to smaller Qwen3 models?
It carries a much larger total parameter budget than the mid-sized and small Qwen3 variants, which generally means more capacity on harder reasoning and coding prompts, but it also costs more per token and needs far more memory to self-host. For simple, high-volume tasks the smaller family members are usually the better cost fit.
How fast is it?
Our benchmark record for this model does not currently include measured output tokens per second or time-to-first-token values. Because the same weights are served by several providers with different hardware and quantization choices, throughput and latency are best compared per provider.
Can I run it on my own hardware?
Qwen3 models are distributed by Alibaba for self-hosting, but a 235B-total-parameter mixture-of-experts model requires enough accelerator memory to hold all experts even though only a subset is active per token. If you are sizing a self-hosted deployment, compare GPU rental costs alongside the hosted API rates in the pricing table on this page.