Activereasoningllm Open Source

Qwen3 235B-A22B

by Alibaba· Released April 2025· Cutoff April 2025

Qwen3 235B-A22B is a Mixture-of-Experts (MoE) large language model with 235B total parameters and 22B activated parameters per token. It is designed for strong reasoning and multilingual capabilities, supporting 119 languages and 29 programming languages, and is part of Alibaba's Qwen3 series.

Official Site API Docs 🤗 Hugging Face

Input cost

Free (open source)

Output cost

Free (open source)

Context window

131072 tokens

Max output

8192 tokens

Modalities

text

Parameters

235B (22B activated)

License

Apache-2.0

Capabilities

Function CallingCode GenerationStreamingJSON ModeMultilingual SupportReasoningAgentic Capabilities

Best For

Complex reasoning tasks, multilingual applications, and agentic workflows requiring efficient inference with MoE architecture.

Strengths

Strong reasoning performance comparable to much larger dense models
Efficient inference with only 22B activated parameters
Broad multilingual support covering 119 languages
Open source with permissive license (Apache-2.0)

Limitations

Requires significant hardware for full model (235B total parameters)
May not match top-tier dense models on certain benchmarks
Relatively new with limited community ecosystem compared to older models
Context window of 131K tokens is large but not the largest available

Use Cases

Multilingual customer support chatbots

Code generation and debugging across 29 programming languages

Complex reasoning and problem-solving in education

Agentic AI systems with tool use and function calling

Content creation and translation in multiple languages

Research and experimentation in MoE architectures

Enterprise applications requiring cost-effective large-scale inference

Improvements Over Previous Model

Introduced MoE architecture with 235B total parameters and 22B activated, improving efficiency over dense Qwen2.5 models
Extended context window from 32K (Qwen2.5) to 131K tokens
Added support for 119 languages (up from ~30 in Qwen2.5)
Improved reasoning capabilities with chain-of-thought and agentic features
Open source under Apache-2.0 license (Qwen2.5 was also open but with different terms)
Supports 29 programming languages for code generation (expanded from previous versions)

Back to all models

Use Cases

Multilingual customer support chatbots

Code generation and debugging across 29 programming languages

Complex reasoning and problem-solving in education

Agentic AI systems with tool use and function calling

Content creation and translation in multiple languages

Research and experimentation in MoE architectures

Enterprise applications requiring cost-effective large-scale inference

Improvements Over Previous Model

Introduced MoE architecture with 235B total parameters and 22B activated, improving efficiency over dense Qwen2.5 models

Extended context window from 32K (Qwen2.5) to 131K tokens

Added support for 119 languages (up from ~30 in Qwen2.5)

Improved reasoning capabilities with chain-of-thought and agentic features

Open source under Apache-2.0 license (Qwen2.5 was also open but with different terms)

Supports 29 programming languages for code generation (expanded from previous versions)