about
Qwen3 235B A22B is Alibaba's latest-generation dense and mixture-of-experts (MoE) model, representing a fundamental advancement in reasoning and agent capabilities. With 235B total parameters and 22B activated (via 128 experts, 8 active), it uniquely supports seamless switching between thinking mode (for complex mathematical reasoning, code generation, and logical problem-solving) and non-thinking mode (for efficient real-time dialogue). Trained with extensive post-training optimization, Qwen3 surpasses previous generations (QwQ and Qwen2.5) on mathematics, coding, and commonsense reasoning while delivering superior human-preference alignment in creative writing, role-playing, and multi-turn conversations. Native context window of 32,768 tokens extends to 131,072 with YaRN scaling, supporting long-document analysis and multi-turn reasoning sessions.
Licenseapache-2.0
Context window(in thousands)32,768
Use cases for Qwen3 235B A22B
- Hybrid Reasoning Tasks: Seamlessly switch between thinking mode (complex math, logical reasoning, coding) and non-thinking mode (fast dialogue) within a single model.
- Agent-Based Automation: Excel at tool calling and agent capabilities with precise integration of external tools, achieving leading performance among open-source models on complex agent tasks.
- Multilingual Applications: Process and respond in 100+ languages and dialects with strong instruction-following capabilities for translation and cross-language reasoning.
Quality
Arena Elo1422
MMLUN/A
MT BenchN/A
Qwen3 235B A22B surpasses previous generations on mathematical reasoning, code generation, and commonsense logical reasoning benchmarks. In thinking mode, performance exceeds QwQ-32B; in non-thinking mode, it matches or exceeds Qwen2.5 Instruct. Leading performance on complex agent-based tasks among open-source models. Benchmarked on MATH, AIME, reasoning competitions, and real-world agent tasks. Supports dynamic output lengths up to 38,912 tokens for complex problem-solving scenarios.
What's Twitter saying?
- Reasoning breakthrough: Qwen3 uniquely combines thinking and dialogue modes, surpassing specialized reasoning models on math and code. Alibaba's technical report shows MATH benchmark improvements over QwQ-32B. src: x.com
- Open-source leadership: Leading performance on agent capabilities among open-source models, with comprehensive tool-calling framework via Qwen-Agent. src: x.com
- Multilingual powerhouse: 100+ language support with native instruction-following for translation, cross-lingual reasoning, and localized agent deployment. src: x.com
Explore Our LLM Library
Discover the power and diversity of large language models available with Telnyx. Explore the options below to find the perfect model for your project.
| Organization | Model Name | Tasks | Languages Supported | Context Length | Parameters | Model Tier | License |
|---|---|---|---|---|---|---|---|
| deepseek-ai | DeepSeek-R1-Distill-Qwen-14B | text generation | English | 43,000 | 14.8B | medium | deepseek |
| fixie-ai | ultravox-v0_4_1-llama-3_1-8b | audio text-to-text | Multilingual | 8,000 | 8.7B | small | mit |
| gemma-2b-it | text generation | English | 8,192 | 2.5B | small | gemma | |
| gemma-7b-it | text generation | English | 8,192 | 8.5B | small | gemma | |
| meta-llama | Llama-3.3-70B-Instruct | text generation | Multilingual | 99,000 | 70.6B | large | llama3.3 |
| meta-llama | Llama-Guard-3-1B | safety classification | Multilingual | 128,000 | 1.5B | small | llama3.3 |
| meta-llama | Meta-Llama-3.1-70B-Instruct | text generation | Multilingual | 99,000 | 70.6B | large | llama3.1 |
| meta-llama | Meta-Llama-3.1-8B-Instruct | text generation | Multilingual | 131,072 | 8.0B | small | llama3.1 |
| minimaxai | MiniMax-M2.5 | text generation | English | 2,000,000 | 0 | large | minimaxai |
| mistralai | Mistral-7B-Instruct-v0.1 | text generation | English | 8,192 | 7.2B | small | apache-2.0 |
| mistralai | Mistral-7B-Instruct-v0.2 | text generation | English | 32,768 | 7.2B | small | apache-2.0 |
| mistralai | Mixtral-8x7B-Instruct-v0.1 | text generation | Multilingual | 32,768 | 46.7B | medium | apache-2.0 |
| moonshotai | Kimi-K2.5 | text generation | English | 256,000 | 1.0T | large | modified-mit |
| Qwen | Qwen3-235B-A22B | text generation | English | 32,768 | 235.1B | large | apache-2.0 |
| zai-org | GLM-5 | text generation | English | 202,752 | 753.9B | large | mit |
| anthropic | claude-3-7-sonnet-latest | text generation | Multilingual | 200,000 | 0 | large | anthropic |
| anthropic | claude-haiku-4-5 | text generation | Multilingual | 200,000 | 0 | large | anthropic |
| anthropic | claude-opus-4-6 | text generation | Multilingual | 200,000 | 0 | large | anthropic |
| anthropic | claude-sonnet-4-20250514 | text generation | Multilingual | 200,000 | 0 | large | anthropic |
| gemini-2.0-flash | text generation | Multilingual | 1,048,576 | 0 | large | ||
| gemini-2.5-flash | text generation | Multilingual | 1,048,576 | 0 | large | ||
| gemini-2.5-flash-lite | text generation | Multilingual | 1,048,576 | 0 | large | ||
| groq | gpt-oss-120b | text generation | English | 131,072 | 117.0B | large | groq |
| groq | kimi-k2-instruct | text generation | English | 131,072 | 1.0T | large | groq |
| groq | llama-3.3-70b-versatile | text generation | Multilingual | 131,072 | 70.6B | large | llama3.3 |
| groq | llama-4-maverick-17b-128e-instruct | text generation | Multilingual | 1,000,000 | 400.0B | large | llama4 |
| groq | llama-4-scout-17b-16e-instruct | text generation | Multilingual | 128,000 | 109.0B | large | llama4 |
| openai | gpt-3.5-turbo | text generation | Multilingual | 4,096 | 0 | large | openai |
| openai | gpt-4 | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-0125-preview | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-0314 | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-0613 | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-1106-preview | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-32k-0314 | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4-turbo-preview | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4.1 | text generation | Multilingual | 1,047,576 | 0 | large | openai |
| openai | gpt-4.1-mini | text generation | Multilingual | 1,047,576 | 0 | large | openai |
| openai | gpt-4o | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-4o-mini | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | gpt-5 | text generation | Multilingual | 400,000 | 0 | large | openai |
| openai | gpt-5-mini | text generation | Multilingual | 400,000 | 0 | large | openai |
| openai | gpt-5.1 | text generation | Multilingual | 400,000 | 0 | large | openai |
| openai | gpt-5.2 | text generation | Multilingual | 400,000 | 0 | large | openai |
| openai | o1-mini | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | o1-preview | text generation | Multilingual | 128,000 | 0 | large | openai |
| openai | o3-mini | text generation | Multilingual | 200,000 | 0 | large | openai |
| xai-org | grok-2 | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-2-latest | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3 | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-beta | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-fast | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-fast-beta | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-fast-latest | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-latest | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-mini | text generation | Multilingual | 131,072 | 0 | large | xai |
| xai-org | grok-3-mini-fast | text generation | Multilingual | 131,072 | 0 | large | xai |
TRY IT OUT
Chat with an LLM
Powered by our own GPU infrastructure, select a large language model, add a prompt, and chat away. For unlimited chats, sign up for a free account on our Mission Control Portal here.
RESOURCES
Get started
Check out our helpful tools to help get you started.
faqs
What makes Qwen3 235B A22B different from other reasoning models?
Qwen3 uniquely supports dynamic switching between thinking and non-thinking modes within a single model, eliminating the need for separate inference paths. This hybrid architecture allows complex reasoning (math, coding) without sacrificing dialogue efficiency. Traditional reasoning models like o1 operate in thinking-only mode; Qwen3's toggle enables cost-optimized deployment. Visit our comparison guide to understand architectural trade-offs.
How does the mixture-of-experts architecture improve performance?
With 235B parameters but only 22B activated per token, Qwen3 achieves superior quality-to-cost ratios compared to dense models. The 128-expert, 8-active configuration enables specialized task expertise while maintaining training efficiency. For technical deep-dive on MoE scaling, see here.
Can Qwen3 handle long documents?
Yes. Natively supports 32,768 tokens; extends to 131,072 with YaRN rope-scaling (4x factor). Recommended for legal contracts, research papers, multi-turn agent workflows. Deployment via vLLM or SGLang with YaRN configuration enabled. Reference for deployment options.
What languages does Qwen3 support?
100+ languages and dialects with strong instruction-following for translation and cross-language reasoning. Multilingual agent capabilities enable localized tool calling and regional deployment. Validated on instruction-following across major language families. See here for multilingual voice-AI applications.
How do I use Qwen3 in agent workflows?
Deploy via Qwen-Agent framework (built-in tool calling), standard OpenAI-compatible APIs (vLLM, SGLang), or custom integrations. MCP (Model Context Protocol) support enables seamless tool binding. Recommended output length: 32,768 tokens for standard tasks, 38,912 for complex problem-solving. Integration examples available at https://developers.telnyx.com.
Is Qwen3 production-ready?
Yes. Designed for enterprise deployment with SLA support via Telnyx Inference. Sampling parameters: thinking mode (Temperature 0.6, TopP 0.95), non-thinking (Temperature 0.7, TopP 0.8). Presence_penalty (0-2) reduces repetitions. Validate on your workload before production rollout. Full deployment guide for compliance-sensitive applications.