Alibaba's flagship Qwen3 model (235B total, 22B active, 128 experts, 8 active per token MoE). Dual thinking/non-thinking mode, strong reasoning, tools, and 100+ language support.
Qwen3-235B-A22B is Alibaba Qwen's open-weight flagship model, released on April 28, 2025 as part of the Qwen3 family. It is a Mixture-of-Experts (MoE) architecture with 235 billion total parameters and 22 billion activated per forward pass — giving it the reasoning depth of a frontier large model at the inference cost of a ~22B dense model. The architecture uses 94 transformer layers, 128 total experts with 8 activated per token, GQA attention (64Q / 4KV heads), and was pretrained on roughly 36 trillion tokens covering 119 languages and dialects. A key architectural innovation is the unified hybrid thinking system: a single model seamlessly switches between thinking mode (step-by-step chain-of-thought) and non-thinking mode (direct, low-latency responses) via the enable_thinking flag or /think / /no_think user-turn prefixes.
For TheRouter users, Qwen3-235B is the recommended choice when quality ceiling matters most: frontier-level math and coding reasoning, complex agentic workflows, and multilingual tasks across 119 languages. It matches or exceeds models such as DeepSeek-R1, o1, and Gemini-2.5-Pro on key benchmarks — notably 85.7 on AIME'24, 81.5 on AIME'25, and 70.7 on LiveCodeBench v5. Because only 22B parameters are active per inference step, it is significantly more cost-efficient than equivalently-sized dense models at this quality level. The non-thinking mode delivers Qwen2.5-Instruct-class speed and conciseness; the thinking mode adds deep reasoning for hard problems without requiring a separate model.
| Type | Rate |
|---|---|
| Input | $0.756 / 1M tokens |
| Output | $3.02 / 1M tokens |
| Release date | 2025-04-28qwenlm.github.io ↗ | verified |
| Architecture | Mixture-of-Experts (MoE); 235B total parameters, 22B activated per token; 94 layers; 128 experts / 8 activated; GQA (64Q / 4KV heads)huggingface.co ↗ | verified |
| Native context length | 32,768 tokens natively; extendable to 131,072 tokens with YaRNhuggingface.co ↗ | verified |
| Thinking modes | Hybrid: thinking mode (chain-of-thought, enable_thinking=True) and non-thinking mode (enable_thinking=False); switchable per turn via /think and /no_think prefixeshuggingface.co ↗ | verified |
| Pretraining data | ~36 trillion tokens; 119 languages and dialects; includes synthetic math/code data from Qwen2.5-Math and Qwen2.5-Coderqwenlm.github.io ↗ | verified |
| Multilingual support | 119 languages and dialectsqwenlm.github.io ↗ | verified |
| License | Apache 2.0qwenlm.github.io ↗ | verified |
| Training cutoff | Not publicly disclosed for the base release; the -2507 variant targets July 2025 knowledgewww.mindstudio.ai ↗ | to verify |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AIME 2024 Thinking mode | 85.7% | arxiv.org ↗ | |
AIME 2025 Thinking mode | 81.5% | arxiv.org ↗ | |
LiveCodeBench v5 Thinking mode | 70.7% | arxiv.org ↗ | |
CodeForces Thinking mode | 2056rating | arxiv.org ↗ | |
BFCL v3 | 70.8% | arxiv.org ↗ |
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/qwen3-235b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-235b",
"messages": [
{"role": "user", "content": "/no_think Summarise the key changes in this diff: ..."}
]
}'The Qwen Team released the full technical report covering the Qwen3 family's pretraining methodology (36T tokens, 3-stage), hybrid thinking architecture, post-training pipeline, and comprehensive benchmark results showing Qwen3-235B-A22B achieving 85.7 AIME'24, 81.5 AIME'25, 70.7 LiveCodeBench v5, and 70.8 BFCL v3 in thinking mode.
Alibaba Qwen released Qwen3-235B-A22B and the full Qwen3 family on Hugging Face, ModelScope, and Kaggle under Apache 2.0, making the largest open-weight hybrid-thinking MoE model publicly available. The release included both pretrained base and post-trained instruct versions, alongside six dense models from 0.6B to 32B and the smaller Qwen3-30B-A3B MoE.
Thinking mode (/think prefix or enable_thinking=True) causes the model to generate a full chain-of-thought reasoning trace inside <think>...</think> tags before the final answer. This significantly improves accuracy on hard math, coding, and multi-step reasoning at the cost of higher token count, latency, and cost. Non-thinking mode (/no_think prefix) produces a direct, concise answer — faster and cheaper, appropriate for conversational, creative, and simpler task types. You can switch mode per turn in the same conversation.
Qwen3-235B-A22B achieves competitive or better scores than DeepSeek-R1 and o1 on math and coding benchmarks (85.7 AIME'24, 81.5 AIME'25, 70.7 LiveCodeBench v5) in thinking mode. The Qwen team positioned it as directly comparable to Grok-3 and Gemini-2.5-Pro as well. The key advantage over closed models is Apache 2.0 open-weight access and the hybrid thinking switch, which allows the same model to serve fast non-reasoning queries without a separate deployment.
Yes. Set your client's base URL to https://api.therouter.ai/v1, configure your TheRouter API key, and set the model to qwen/qwen3-235b. Streaming and tool calling are supported. To activate thinking mode, prefix user messages with /think; for direct responses use /no_think. Clients that send plain text with no prefix will use the model's default mode — check TheRouter docs for the current default.
Qwen3-235B is a Mixture-of-Experts model: its feed-forward layers contain 128 expert sub-networks, of which only 8 are selected (via a learned routing mechanism) for each token during inference. The remaining 120 experts are dormant for that token. This means the compute cost per token is equivalent to roughly a 22B dense model, while the model's total capacity — spread across all 128 experts — reaches 235B parameters. The result is frontier reasoning quality at dramatically lower inference cost compared to a hypothetical 235B dense model.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Architecture | huggingface.co ↗ | 2026-06-09 | verified |
| Native context length | huggingface.co ↗ | 2026-06-09 | verified |
| Thinking modes | huggingface.co ↗ | 2026-06-09 | verified |
| Pretraining data | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Multilingual support | qwenlm.github.io ↗ | 2026-06-09 | verified |
| License | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Training cutoff | www.mindstudio.ai ↗ | 2026-06-09 | to verify |
| AIME 2024 | arxiv.org ↗ | 2026-06-09 | verified |
| AIME 2025 | arxiv.org ↗ | 2026-06-09 | verified |
| LiveCodeBench v5 | arxiv.org ↗ | 2026-06-09 | verified |
| CodeForces | arxiv.org ↗ | 2026-06-09 | verified |
| BFCL v3 | arxiv.org ↗ | 2026-06-09 | verified |
| Qwen3 technical report published on arXiv (2505.09388) | arxiv.org ↗ | 2026-06-09 | verified |
| Qwen3-235B-A22B released under Apache 2.0 | qwenlm.github.io ↗ | 2026-06-09 | verified |
| What is the difference between qwen/qwen3-235b thinking mode and non-thinking mode on TheRouter? | huggingface.co ↗ | 2026-06-09 | to verify |
| How does Qwen3-235B compare to DeepSeek-R1 and o1? | arxiv.org ↗ | 2026-06-09 | to verify |
| Can I use Qwen3-235B from Cursor, Cline, or an OpenAI-compatible SDK? | huggingface.co ↗ | 2026-06-09 | to verify |
| Why does Qwen3-235B only activate 22B parameters if it has 235B total? | huggingface.co ↗ | 2026-06-09 | to verify |