Qwen/Qwen3-32B¶
Model Information¶
Qwen/Qwen3-32B is a multilingual, instruction-tuned large language model developed by Alibaba Cloud. It is part of the Qwen3 series, designed to rival top-tier open and commercial LLMs in reasoning, multilingual QA, coding, and alignment. With 32 billion parameters, it balances strong performance with manageable deployment costs.
- Model Developer: Alibaba Cloud (Qwen Team)
- Model Release Date: May 2024
- Supported Languages: English, Chinese, French, Spanish, German, Japanese, Korean, Portuguese, and other major languages.
Model Architecture¶
Qwen/Qwen3-32B is a dense, decoder-only transformer model designed for instruction-following and multilingual tasks.
Key Architecture Details:
- Model Type: Decoder-only transformer
- Parameters: 32B
- Context Length: Up to 128K tokens
-
Training Strategy:
- Pretraining on a diverse multilingual and code corpus
- Instruction fine-tuning using SFT and preference-aligned data
- Safety alignment and steering tuning for compliant outputs
- Tokenizer: Optimized multilingual tokenizer from Qwen3 series
- Capabilities:
- Strong few-shot and zero-shot generalization
- Robust code generation and math reasoning
- Steerable behavior through system prompts
Benchmark Scores¶
| Category | Benchmark | Shots | Metric | Qwen3-32B |
|---|---|---|---|---|
| General | MMLU (CoT) | 0 | Acc. (avg) | 88.0 |
| MMLU Pro (CoT) | 5 | Acc. (avg) | 60.1 | |
| Steerability | IFEval | – | – | 92.1 |
| Reasoning | GPQA Diamond (CoT) | 0 | Accuracy | 47.3 |
| Code | HumanEval | 0 | Pass@1 | 85.4 |
| MBPP EvalPlus (base) | 0 | Pass@1 | 85.9 | |
| Math | MATH (CoT) | 0 | Sympy Score | 59.4 |
| Tool Use | BFCL v2 | 0 | AST Macro Avg. | 80.3 |
| Multilingual | MGSM | 0 | EM (exact match) | 77.6 |
Qwen3-32B competes closely with leading models like GPT-4o and Gemini Flash, offering robust multilingual reasoning, math, and tool use performance.