Qwen3-235B-A22B: specs, benchmarks, and how to run it on Lyceum

Get started: call Qwen3-235B-A22B on Lyceum

Integrate Qwen3-235B-A22B into your application using the standard OpenAI SDK. Lyceum Technology provides a drop-in replacement API, meaning you only need to update your base URL and API key to start routing requests to our European data centers.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="Qwen/Qwen3-235B-A22B-Instruct-2507",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Qwen3-235B-A22B

When you deploy Qwen3-235B-A22B on Lyceum Technology, you benefit from transparent, per-token pricing with no hidden base fees or minimum commitments. The model is hosted in our eu-north1 region, so processing for this model runs in European data centres, with no training on customer data and no data retained after processing.

This model is served through Lyceum's Serverless Inference, the pay-per-token, OpenAI-compatible endpoint for open-weight models. There are no service tiers: every model on it is billed the same way. The pricing for the Qwen/Qwen3-235B-A22B-Instruct-2507 endpoint is set at $0.20 per million input tokens and $0.60 per million output tokens. Because billing is per token with no base fee, these rates stay predictable as your traffic changes. You only pay for the exact number of tokens you process, making it highly cost-effective for both bursty workloads and sustained production traffic.

For engineers migrating from existing cloud providers, the transition requires zero code refactoring. The endpoint fully supports standard parameters like temperature, max_tokens, and stream, allowing you to maintain your existing application logic while moving to an environment that supports your GDPR compliance.

What Qwen3-235B-A22B is good at

Efficient Mixture-of-Experts architecture

Qwen3-235B-A22B represents a significant architectural leap for the Alibaba Cloud Qwen team. It utilizes a Mixture-of-Experts (MoE) design, housing a massive 235 billion total parameters. However, during any given forward pass, the model activates only 8 of its 128 experts, utilizing 22 billion parameters per token. This sparse activation strategy allows the model to achieve the reasoning depth and knowledge retention of a massive dense model while maintaining the inference speed and computational efficiency of a much smaller 22B parameter model. For engineering teams, this translates to faster time-to-first-token (TTFT) and higher throughput in production.

Massive 256K context window

One of the standout capabilities of the Instruct-2507 checkpoint is its native support for a 262,144-token context window. This massive capacity allows the model to process hundreds of pages of text, extensive codebases, or large datasets in a single prompt. It is particularly effective for document parsing, retrieval-augmented generation (RAG) pipelines, and summarizing long-form enterprise data without requiring aggressive chunking strategies.

Multilingual proficiency and coding

Qwen3-235B-A22B excels in multilingual environments. The Qwen3 Technical Report states that Qwen3 expands multilingual support from 29 to 119 languages and dialects. Its training data heavily emphasizes cross-lingual alignment, making it an excellent choice for global applications requiring translation or localized instruction following. Furthermore, the model demonstrates exceptional proficiency in software engineering tasks. It reliably generates, debugs, and refactors code across multiple programming languages, and its robust tool-usage capabilities allow it to smoothly integrate with external APIs and function-calling frameworks in complex agentic workflows. The smaller member of the family is covered in our notes on Qwen3-30B-A3B and how to run it.

Limitations and trade-offs

Where Qwen3-235B-A22B falls short

While Qwen3-235B-A22B is a highly capable foundation model, it is important to understand the specific trade-offs of the Instruct-2507 checkpoint. Unlike the "Thinking" variants in the Qwen3 family, this specific model is optimized for rapid, non-thinking text generation. It does not output <think> blocks or engage in extended chain-of-thought reasoning before returning an answer. Consequently, while it is exceptionally fast for general-purpose dialogue, data extraction, and standard coding tasks, it may struggle with highly complex mathematical proofs or multi-step logical puzzles that require deep, iterative reasoning.

Additionally, because it relies on a sparse Mixture-of-Experts architecture, its knowledge retrieval on highly niche or obscure topics can sometimes fall short compared to massive dense models. While the 235B total parameter count provides a vast knowledge base, the activation of only 22B parameters per token means that the model can occasionally hallucinate specific factual details when pushed outside of its primary training distribution.

Managing the full 256K context window requires careful attention to prompt engineering. While the model can ingest massive amounts of text, performance and recall accuracy can degrade slightly when critical information is buried in the middle of an extremely long prompt - a common phenomenon known as the "lost in the middle" effect. Engineering teams should still employ efficient retrieval-augmented generation (RAG) techniques rather than relying solely on the maximum context window for large-scale knowledge retrieval.

Benchmarks and how it compares

Qwen3-235B-A22B benchmark results

The Qwen3-235B-A22B-Instruct-2507 model has been rigorously evaluated across industry-standard benchmarks, demonstrating performance that rivals top-tier proprietary models and significantly outpaces previous open-weight generations. According to the official Hugging Face model card, the model excels in general knowledge, coding, and instruction following.

Source: the Qwen3-235B-A22B-Instruct-2507 model card published by the Alibaba Cloud Qwen team, read 3 August 2026.

Compared to its predecessor, Qwen2.5-72B-Instruct, the new Qwen3-235B-A22B model offers a substantial leap in capability. While the 72B model is a dense architecture, the 235B MoE model provides a much wider breadth of knowledge and superior multilingual support while maintaining similar inference speeds due to its 22B active parameter count.

Against current proprietary models, Qwen3-235B-A22B holds its ground remarkably well. It frequently matches or exceeds the performance of models in its weight class on coding and tool-usage tasks. For European engineering teams, this means you can reach proprietary-level performance with an open-weight model served from European data centres, and keep the freedom to move the weights elsewhere that a closed model does not offer.

Using it in production

Production configuration for Qwen3-235B-A22B

Deploying Qwen3-235B-A22B in a production environment requires understanding how to optimize its parameters and manage costs effectively. On Lyceum Technology, this model is served through Serverless Inference, billed per token with no base fee and no minimum commitment.

The model supports a native context window of 262,144 tokens. When processing large documents, we highly recommend enabling the stream=True parameter in your API calls. Streaming reduces perceived latency for end-users, as the model's 22B active parameters can begin returning tokens almost instantly, even when processing extensive input contexts. For tasks requiring strict formatting, such as JSON extraction or function calling, setting a lower temperature (e.g., 0.1 or 0.2) will yield the most deterministic and reliable outputs.

Cost management is straightforward with Lyceum's per-token billing model. The Qwen/Qwen3-235B-A22B-Instruct-2507 endpoint is priced at $0.20 per million input tokens and $0.60 per million output tokens. For example, if your application processes a 50,000-token legal contract and generates a 1,000-token summary, the input cost would be $0.01, and the output cost would be $0.0006, resulting in a total transaction cost of $0.0106.

Because this model is served from the eu-north1 region, inference requests for it are processed in European data centres. That keeps latency low for European users and supports your own adherence to regional data protection rules.

Running Qwen3-235B-A22B in European data centres

Why run Qwen3-235B-A22B on Lyceum

For European AI startups and enterprise engineering teams, where inference actually runs is a live regulatory question. Qwen3-235B-A22B on Lyceum Technology is served from the eu-north1 region, so processing for this model happens in European data centres, with no training on customer data and no prompts or outputs retained after processing. That supports your own GDPR compliance on a self-asserted basis rather than conferring it. Lyceum holds no ISO 27001 or SOC 2 certificate today and states no EU AI Act conformity position; data centre operators hold ISO certifications at facility level.

Lyceum runs in European data centres in Spain, Paris and the Nordics. The serving stack is open: vLLM, NVIDIA Dynamo and TensorRT-LLM, which is what keeps throughput high and cost per token low. Serverless Inference is billed per token with no minimum commitment and no egress fees for data transfer, and GPU VMs, if you need dedicated hardware instead, are billed per second with no base fee.

Integration into your stack is straightforward. Our Serverless Inference endpoint is OpenAI-compatible and works as a drop-in replacement for the OpenAI SDK. You can migrate your existing applications to Qwen3-235B-A22B in minutes by updating your base URL and API key. Scale a high-traffic customer support agent or process massive batches of documents: Lyceum provides the reliability and speed necessary to run production AI workloads, while supporting, but not conferring, your compliance.

Sources

[1] Qwen3-235B-A22B-Instruct-2507 - Hugging Face; [2] Qwen3-235B-A22B - Hugging Face; [3] arXiv: Qwen3 Technical Report