Qwen3 dense 32B model. Excellent reasoning and coding at moderate size with thinking mode support.
Qwen3 32B is Alibaba Qwen's dense 32.8B-parameter open-weight reasoning model. It sits between the smaller Qwen3 dense models and the larger Qwen3 MoE flagships: still practical for controlled serving, but large enough for serious coding, math, multilingual chat, and tool-using agent workloads.
TheRouter exposes it as qwen/qwen3-32b with OpenAI-compatible chat completions, 131K context routing, tools, JSON mode, and a reasoning parameter. Use it when you want a single Qwen model that can switch between fast non-thinking responses and deeper thinking-mode reasoning without moving to a much larger flagship.
| Type | Rate |
|---|---|
| Input | $0.1728 / 1M tokens |
| Output | $0.6912 / 1M tokens |
| Architecture | Dense causal language model (all 32.8B parameters active per token)huggingface.co ↗ | verified |
| Parameters | 32.8B total (31.2B non-embedding)huggingface.co ↗ | verified |
| Layers | 64 transformer layershuggingface.co ↗ | verified |
| Attention heads (GQA) | 64 query heads / 8 KV heads (GQA)huggingface.co ↗ | verified |
| Native context window | 32,768 tokens (native); 131,072 tokens with YaRN positional interpolationhuggingface.co ↗ | verified |
| Thinking mode | Hybrid: single checkpoint supports both thinking (chain-of-thought in <think> blocks) and non-thinking mode; switch via enable_thinking flag or /think / /no_think tokenhuggingface.co ↗ | verified |
| Language support | 119 languages and dialects (Indo-European, Sino-Tibetan, Afro-Asiatic, Austronesian, Dravidian, Turkic, Tai-Kadai, Uralic, and more)qwenlm.github.io ↗ | verified |
| License | Apache 2.0huggingface.co ↗ | verified |
| Pretraining data | ~36 trillion tokens across 119 languages (approx. 2× Qwen2.5's 18T); includes web, PDF extraction, synthetic math/code dataqwenlm.github.io ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AIME 2024 Thinking mode; reported in the Qwen3 official blog as part of the Qwen3-32B evaluation table. | 85.7%% | qwenlm.github.io ↗ | |
MATH-500 Thinking mode; from Qwen3 official blog evaluation table. | 95.5%% | qwenlm.github.io ↗ | |
LiveCodeBench Thinking mode; from Qwen3 official blog evaluation table. | 65.9%% | qwenlm.github.io ↗ | |
BFCL (Berkeley Function-Calling Leaderboard) Tool-calling benchmark; non-thinking mode; from Qwen3 official blog. | 70.8%% | qwenlm.github.io ↗ | |
MMLU-Pro Non-thinking mode; from Qwen3 official blog evaluation table. | 74.9%% | qwenlm.github.io ↗ | |
C-Eval Chinese comprehensive evaluation; from Qwen3 official blog. | 91.8%% | qwenlm.github.io ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "qwen/qwen3-32b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat completions. Omit the reasoning parameter for non-thinking (fast) mode; set reasoning: {} or a budget to enable thinking mode.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-32b",
"messages": [
{"role": "user", "content": "Explain the tradeoffs between dense and MoE architectures for inference."}
],
"temperature": 0.7,
"max_tokens": 2048
}'Alibaba's Qwen3 release includes Qwen3-32B alongside five other dense models and two MoE models, emphasizing hybrid thinking/non-thinking operation, tool use, and multilingual coverage.
Use it as a balanced general reasoning model: coding help, math-heavy analysis, multilingual support, JSON/tool workflows, and medium-stakes agents where a dense 32B model is enough and a larger flagship would waste margin.
Yes. Qwen3 was launched with hybrid thinking and non-thinking modes. On TheRouter, use qwen/qwen3-32b through the OpenAI-compatible endpoint and pass the reasoning parameter when you want deeper reasoning behavior.
The official model card lists 32,768 native tokens and 131,072 tokens with YaRN. TheRouter catalog exposes the routed context length as 131K tokens for qwen/qwen3-32b.
Qwen describes Qwen3-32B as an open-weight dense model released under Apache 2.0. Treat the license as source-level permission; production API usage still depends on the provider route and TheRouter billing path you choose.
Use qwen/qwen3-32b when the task mixes code with general reasoning, math, multilingual product context, or tool use. Use qwen/qwen3-coder-30b when the workload is mostly code generation, autocomplete, tests, and lower-latency coding throughput.
| source | URL | retrieved | |
|---|---|---|---|
| Architecture | huggingface.co ↗ | 2026-06-08 | verified |
| Parameters | huggingface.co ↗ | 2026-06-08 | verified |
| Layers | huggingface.co ↗ | 2026-06-08 | verified |
| Attention heads (GQA) | huggingface.co ↗ | 2026-06-08 | verified |
| Native context window | huggingface.co ↗ | 2026-06-08 | verified |
| Thinking mode | huggingface.co ↗ | 2026-06-08 | verified |
| Language support | qwenlm.github.io ↗ | 2026-06-08 | verified |
| License | huggingface.co ↗ | 2026-06-08 | verified |
| Pretraining data | qwenlm.github.io ↗ | 2026-06-08 | verified |
| AIME 2024 | qwenlm.github.io ↗ | 2026-06-08 | verified |
| MATH-500 | qwenlm.github.io ↗ | 2026-06-08 | verified |
| LiveCodeBench | qwenlm.github.io ↗ | 2026-06-08 | verified |
| BFCL (Berkeley Function-Calling Leaderboard) | qwenlm.github.io ↗ | 2026-06-08 | verified |
| MMLU-Pro | qwenlm.github.io ↗ | 2026-06-08 | verified |
| C-Eval | qwenlm.github.io ↗ | 2026-06-08 | verified |
| Qwen3 launches dense 32B open-weight model with hybrid thinking modes | Qwen3 announcement ↗ | 2026-06-08 | verified |