Qwen3-30B-A3B: specs, benchmarks, and how to run it on Lyceum

Get started: call Qwen3-30B-A3B on Lyceum

Connect directly to the Lyceum Serverless Inference API. Because the platform is fully OpenAI-compatible, integrating this highly efficient Mixture-of-Experts (MoE) model requires zero architectural changes to your existing application. You only need to update your client's base URL and provide your Lyceum API key.

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/Qwen3-30B-A3B-Instruct-2507",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Qwen3-30B-A3B

When you deploy this model on Lyceum, you benefit from our transparent, per-token billing model with no minimum commitments or idle costs. Qwen3-30B-A3B is served through Serverless Inference, which is optimized for cost-efficient, high-throughput workloads where latency and economy are paramount. The pricing is set at $0.10 per million input tokens and $0.30 per million output tokens.

This model is hosted in our eu-north1 region. For European AI teams and enterprises with strict data residency requirements, this ensures that all inference data is processed entirely within the European Union. Because inference for this model runs in European data centres rather than on US-based hyperscalers, processing stays GDPR-compliant while you take advantage of the model's speed and instruction-following capabilities.

What Qwen3-30B-A3B is good at

Efficient MoE Architecture

Qwen3-30B-A3B represents a significant leap in architectural efficiency for open-weight models. It utilizes a highly optimized Mixture-of-Experts (MoE) design. While the model contains 30.5 billion total parameters, it only activates 3.3 billion parameters during any single forward pass. The architecture consists of 128 distinct expert networks, with the routing mechanism selecting exactly 8 active experts per token. This sparse activation allows the model to deliver the nuanced understanding of a 30B-class model while operating at the inference speed and computational cost of a much smaller 3B model. All 128 experts stay resident in memory, however, so the weight footprint tracks the full 30.5 billion parameters and hardware should never be sized from the 3.3 billion activated figure.

Instruction Following and Coding

The specific variant served on Lyceum, Qwen3-30B-A3B-Instruct-2507, is fine-tuned heavily for instruction following, tool usage, and coding tasks. According to the official Qwen release documentation, this model demonstrates substantial gains in long-tail knowledge coverage across multiple languages and excels at complex agentic workflows. It supports over 100 languages and dialects, making it highly versatile for global applications. We cover a model of similar size from another family in our notes on Nemotron-3-Nano-30B and how to run it.

Speed and Cost-Efficiency

Because it only activates 3.3 billion parameters, Qwen3-30B-A3B is exceptionally fast. It is designed to maximize tokens per second, making it an ideal choice for latency-sensitive applications like real-time chatbots, large-scale document parsing, and high-volume data extraction. The model's efficiency allows engineering teams to process massive datasets without incurring the prohibitive costs associated with dense frontier models. For teams transitioning off expensive hyperscaler credits, this MoE architecture provides a sustainable path for scaling production AI workloads.

Limitations and trade-offs

Where Qwen3-30B-A3B falls short

While Qwen3-30B-A3B is a highly capable and efficient model, it is important to understand its architectural trade-offs to determine if it fits your specific use case. The most significant limitation of the Qwen3-30B-A3B-Instruct-2507 variant is that it operates exclusively in non-thinking mode. Unlike the base Qwen3 models or the larger Qwen3-235B which can dynamically switch into a dedicated thinking mode for complex, multi-step logical reasoning, this specific instruct variant does not generate think blocks. It is not the best choice for advanced mathematical proofs, deep scientific reasoning, or tasks that require extensive test-time compute to arrive at an answer.

Because the model only activates 3.3 billion parameters per token, its raw factual recall capacity is inherently lower than that of much larger models such as the dense Llama 3.3 70B or the flagship Qwen3-235B-A22B. While it performs exceptionally well on standard benchmarks, it may occasionally hallucinate obscure facts or struggle with highly niche domain knowledge unless augmented with Retrieval-Augmented Generation (RAG) or external tool use.

While the model supports a massive context window natively, processing extremely long contexts, approaching 131k to 262k tokens, will still consume significant VRAM and compute resources, potentially offsetting the speed advantages of its sparse MoE architecture. For workloads requiring deep reasoning over massive documents, a larger model with dedicated reasoning capabilities might yield more accurate results, albeit at a higher latency and cost.

Benchmarks and how it compares

Qwen3-30B-A3B benchmark results

Despite its small active parameter footprint, Qwen3-30B-A3B delivers highly competitive performance across industry-standard evaluations. The Qwen team's official benchmarks demonstrate that this MoE model frequently outperforms older, larger dense models and even specialized reasoning models in specific categories. It outcompetes QwQ-32B, a model with ten times the active parameters, on several key metrics.

Below are the published benchmark results for the Qwen3-30B-A3B-Instruct-2507 non-thinking variant, as reported in the official Hugging Face model card:

Source: Official Qwen3-30B-A3B-Instruct-2507 Hugging Face Model Card .

When compared to its much larger sibling, the Qwen3-235B operating in non-thinking mode, the 30B-A3B Instruct variant shows significant optimization. It actually scores higher on AIME25 (61.3 vs 24.7) and GPQA (70.4 vs 62.9) than the base 235B model operating without its thinking mode. This highlights the effectiveness of the 2507 instruction-tuning process, making it an effective tool for developers who need high performance without the latency overhead of a 200B parameter model.

Using it in production

Production configuration for Qwen3-30B-A3B

When deploying Qwen3-30B-A3B in a production environment, understanding its configuration parameters and cost structure is essential for optimizing your application. The model natively supports a massive context length of up to 262,144 tokens. This makes it highly suitable for processing large codebases, analyzing extensive financial reports, or handling long-running conversational agents without losing context. However, developers should monitor their input token counts, as maximizing the context window will impact overall request latency.

On Lyceum, Qwen3-30B-A3B is served through Serverless Inference and billed per token, with no service tiers to choose between. Per-token billing suits workloads that prioritize high throughput and cost-efficiency, making it ideal for tasks like batch document processing, real-time API serving, and high-volume text extraction.

The per-token pricing model ensures you only pay for exact usage, with no base fees or idle costs. At $0.10 per million input tokens and $0.30 per million output tokens, the economics are highly favorable for scale. For example, if your application processes 5 million input tokens and generates 2 million output tokens daily, your total daily cost would be $1.10, calculated as $0.50 for input and $0.60 for output. This predictable, low-cost structure is a major advantage for AI startups and scale-ups looking to move away from expensive, dedicated hyperscaler instances. Because Lyceum does not charge egress fees, you can move large datasets in and out of our S3-compatible storage without incurring hidden network transfer costs.

Running Qwen3-30B-A3B on EU-sovereign infrastructure

Why run Qwen3-30B-A3B on Lyceum

For European AI teams, compliance and data sovereignty are critical requirements. By running Qwen3-30B-A3B on Lyceum, your workloads are executed entirely on infrastructure that Lyceum describes as EU-sovereign. The model is hosted in our eu-north1 region, ensuring that all data processing and storage remain strictly within European borders. This supports GDPR-compliant processing in European data centres. Lyceum holds no ISO 27001 certificate and states no EU AI Act conformity position; data centre operators hold ISO certifications at facility level.

Lyceum provides an open-stack transparency advantage. Unlike competitors who rely on black-box proprietary engines, our inference stack leverages open-source technologies like vLLM and NVIDIA Dynamo. This ensures that you maintain complete customer portability and avoid vendor lock-in. You get the performance benefits of advanced inference orchestration without sacrificing control over your deployment architecture.

Integrating the model requires minimal effort thanks to our fully OpenAI-compatible API. Your engineering team can switch from existing providers to Lyceum by changing two lines of code, the base URL and the API key. Combined with a serving stack tuned for high-throughput MoE workloads, Lyceum delivers Qwen3-30B-A3B at a competitive price. You benefit from per-token billing, scale-to-zero capabilities, and capacity in Lyceum's European data centres in Spain, France and the Nordics.

Sources

[1] Qwen3-30B-A3B-Instruct-2507 Model Card; [2] Qwen3-30B-A3B Base Model Card; [3] Qwen3: Think Deeper, Act Faster; [4] Qwen3 Technical Report (arXiv)