# Qwen3 32B is available on Lambda's Inference API

May 29, 2025• 7 min read

[Alibaba](https://www.alibaba.com/) released one of the most advanced models to date, [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B), which can now be integrated into your applications with [Lambda](https://lambda.ai/)’s [Inference API](https://lambda.ai/inference). With its dense model architecture, hybrid reasoning, multilingual support, and agentic capabilities, [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) is designed for complex tasks that would otherwise require human intervention.

## **What is the model architecture of Qwen3 32B?**

[**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) is a dense model, a term used to describe neural networks whose neurons are structured in layers that are interconnected with each neuron in the next layer. These models can easily learn about interactions between input features and non-linear functions because of the large volume of data that can be embedded in its parameters. As the name suggests, [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) is an LLM with 32 billion total parameters.

[**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) competes with larger models like [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1), [xAI](https://x.ai/)’s [Grok-3](https://x.ai/news/grok-3), [OpenAI](http://openai.com/)’s [o1](https://openai.com/o1/) and [o3-mini](https://openai.com/index/openai-o3-mini/) on benchmarks for coding, math, and logic. You can use our guides to reproduce evaluations for the coding benchmark [LiveCodeBench](https://github.com/LambdaLabsML/LiveCodeBench/blob/main/Guide.md) and [GPQA](https://github.com/LambdaLabsML/lm-evaluation-harness/blob/main/Guide.md), a graduate-level Google-proof Q&A benchmark.

## **What are the capabilities of Qwen3 32B?**

With its STEM and logical reasoning proficiency, [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) has two problem-solving modes, commonly known as _hybrid reasoning_. _Thinking mode_, which is enabled by default, helps a model process the inquiry and reason sequentially before responding. In _non-thinking mode_, the model answers simple questions instantly. Developers can switch modes by changing the open-source code, and users can enter /think or /nothink commands to toggle between the two options.

Not only is [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) capable of hybrid reasoning, but it also excels in code generation and tasks that previously required human-performance – like creative writing, role-playing, instruction following, and multi-turn dialogue. Additionally, [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) supports 119 of the most widely-spoken languages and nuanced dialects.

[**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) can execute agentic actions, enabling developers to call their tools of choice and to modify [_Model Context Protocol_](https://www.anthropic.com/news/model-context-protocol) (MCP) configuration files, which are standards that define how LLMs integrate with data sources or third-party applications. Here is an example of a Python script that lets you define functions and deploy agents with access to [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) provided by [Lambda](https://lambda.ai/)’s [Inference API](https://lambda.ai/inference).

```
from qwen_agent.agents import Assistant

# Define LLM
llm_cfg = {
    'model': 'Qwen3-32B',

# Use Lambda's Inference API:
    'model_server': 'https://api.lambdalabs.com/api/v1/inference',  # Replace with the actual endpoint URL
    'api_key': 'YOUR_LAMBDA_API_KEY',  # Replace with your actual API key

# Other parameters:
    # 'generate_cfg': {
    #         # Add: When the response content is `<think>this is the thought</think>this is the answer;
    #         # Do not add: When the response has been separated by reasoning_content and content.
    #         'thought_in_content': True,
    #     },
}

# Define Tools
tools = [\
    {'mcpServers': {  # You can specify the MCP configuration file\
            'time': {\
                'command': 'uvx',\
                'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']\
            },\
            "fetch": {\
                "command": "uvx",\
                "args": ["mcp-server-fetch"]\
            }\
        }\
    },\
  'code_interpreter',  # Built-in tools\
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]
for responses in bot.run(messages=messages):
    pass
print(responses)
```

## **How was Qwen3 32B trained?**

[**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) was pre-trained with 36 trillion tokens from online sources and PDF documents. The legacy multi-modal [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL) model extracted and improved the text quality of these files, while [Qwen2.5-Math](https://huggingface.co/collections/Qwen/qwen25-math-66eaa240a1b7d5ee65f1da3e) and [Qwen2.5-Coder](https://huggingface.co/collections/Qwen/qwen25-coder-66eaa22e6f99801bf65b0c2f) generated synthetic math and coding data from textbooks, code snippets, and question-answer pairs.

Compared to the [Qwen2.5](https://huggingface.co/collections/Qwen/qwen25-66e81a666513e518adb90d9e) base models, the smaller [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) model delivers similar performance, and therefore operates more efficiently due to the following pre-training processes:

1. [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) was taught basic language skills and general knowledge with 30 trillion tokens at a context length of four thousand tokens.
2. Training datasets were improved with an additional five trillion tokens for STEM, coding, and reasoning tasks.
3. Context lengths were extended to 32 thousand tokens with high-quality, long-context datasets.

Models with hybrid-thinking functions were developed using a four-stage post-training pipeline.

1. In the first stage, models were fine-tuned with long _Chain-of-Thought_ (CoT) data, describing its thought-process before undergoing various tasks and disciplines.
2. The second stage required scaling computational resources for reinforcement learning to enhance the models’ ability to explore knowledge and exploit rules-based rewards.
3. During stage three, the [Qwen team](https://qwenlm.github.io/about/) integrated non-thinking capabilities by fine-tuning models with instruction-tuning and additional long CoT data generated by the thinking model to develop quick, but well-thought responses.
4. Lastly, reinforcement learning was applied to 20 generic tasks to rectify unwanted model behavior and improve instruction and format following as well as agentic actions.

## **How to start testing and using Qwen3 32B**

API calls for [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) are priced at $0.10 per million input tokens and $0.30 per million output tokens. It’s available with an FP8 weight precision level, and context windows can reach up to 41,000 tokens in length. This makes [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) and [Lambda](https://lambda.ai/)’s [Inference API](https://lambda.ai/inference) as cost-efficient as possible, with no rate limits. [**Qwen3-32B**](https://huggingface.co/Qwen/Qwen3-32B) can now be integrated into applications with an [API key](https://lambda.ai/inference). Review our [technical documentation](https://docs.lambda.ai/public-cloud/lambda-inference-api/#using-the-lambda-inference-api) for more details.