Activefrontierllm Open Source

Qwen3 235B

by Alibaba· Released April 2025· Cutoff March 2025

Qwen3 235B is Alibaba's largest and most capable MoE (Mixture-of-Experts) language model, featuring 235 billion total parameters with 35 billion activated per token. It achieves state-of-the-art performance across reasoning, coding, and multilingual tasks, rivaling leading models like DeepSeek-R1 and GPT-4o.

Official Site API Docs 🤗 Hugging Face 📄 Research Paper

Input cost

Free (open source)

Output cost

Free (open source)

Context window

131072 tokens

Max output

8192 tokens

Modalities

text

Parameters

235B

License

Apache-2.0

Capabilities

Function CallingCode GenerationReasoningMultilingual SupportStreamingJSON ModeTool UseAgentic Workflows

Best For

Complex reasoning, advanced coding, and multilingual tasks requiring a frontier-level open-source model.

Strengths

Top-tier reasoning and math performance (AIME 2024: 96.2%, MATH-500: 97.9%)
Excellent coding benchmarks (LiveCodeBench: 79.4%, SWE-bench Verified: 72.6%)
Strong multilingual capabilities across 119 languages
Efficient MoE architecture with only 35B activated parameters per token
Open-source with permissive Apache 2.0 license

Limitations

Very large total parameter count (235B) requires substantial hardware for full deployment
Context window of 131K tokens is smaller than some competitors (e.g., 1M+ tokens)
May not be as optimized for very long-context tasks compared to specialized models
Relatively new model with limited third-party evaluations and ecosystem support

Use Cases

Advanced code generation and debugging

Complex mathematical problem solving

Multilingual translation and content creation

Scientific research and data analysis

Building autonomous AI agents and tool-using systems

Enterprise-level reasoning and decision support

Educational tutoring for STEM subjects

Improvements Over Previous Model

Introduced MoE architecture with 235B total parameters (35B activated), a major shift from dense predecessors like Qwen2.5-72B
Significantly improved reasoning performance: AIME 2024 score of 96.2% vs Qwen2.5-72B's 43.3%
Enhanced coding capabilities: LiveCodeBench 79.4% vs Qwen2.5-72B's 50.0%
Extended context window from 32K tokens (Qwen2.5) to 131K tokens
Added native support for tool use and agentic workflows
Open-sourced under Apache 2.0 license, whereas Qwen2.5 used a custom license

Back to all models

Strengths

Top-tier reasoning and math performance (AIME 2024: 96.2%, MATH-500: 97.9%)

Excellent coding benchmarks (LiveCodeBench: 79.4%, SWE-bench Verified: 72.6%)

Strong multilingual capabilities across 119 languages

Efficient MoE architecture with only 35B activated parameters per token

Open-source with permissive Apache 2.0 license

Limitations

Very large total parameter count (235B) requires substantial hardware for full deployment

Context window of 131K tokens is smaller than some competitors (e.g., 1M+ tokens)

May not be as optimized for very long-context tasks compared to specialized models

Relatively new model with limited third-party evaluations and ecosystem support

Improvements Over Previous Model

Introduced MoE architecture with 235B total parameters (35B activated), a major shift from dense predecessors like Qwen2.5-72B

Significantly improved reasoning performance: AIME 2024 score of 96.2% vs Qwen2.5-72B's 43.3%

Enhanced coding capabilities: LiveCodeBench 79.4% vs Qwen2.5-72B's 50.0%

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

Added native support for tool use and agentic workflows

Open-sourced under Apache 2.0 license, whereas Qwen2.5 used a custom license