Qwen3 14B — balanced mid-range model with strong reasoning at low cost.
Qwen3-14B is a 14.8-billion-parameter dense causal language model from Alibaba's Qwen team — the mid-range entry in the Qwen3 dense lineup between the efficient Qwen3-8B and the larger Qwen3-32B. Built with 40 transformer layers and 40/8 GQA attention heads, it was released under Apache 2.0 in April 2025. A key architectural insight from the Qwen3 technical report: Qwen3-14B's base model performance matches that of Qwen2.5-32B-Base, meaning the Qwen3 generation delivers a full generation-jump in parameter efficiency. Concretely, you get Qwen2.5-32B-class capability for less than half the parameter count and at significantly lower inference cost.
Like all Qwen3 instruct models, Qwen3-14B ships with hybrid thinking as a native feature: a single checkpoint switches between a deep-reasoning thinking mode (step-by-step reasoning in <think> blocks) and a fast non-thinking mode, controlled per-request via enable_thinking or /think / /no_think tokens. This matters for routing: teams can assign the same model ID to both quick completions and harder analytical tasks, differentiating behavior through TheRouter's reasoning parameter rather than managing two separate model deployments. At $0.084 / $0.252 per million input/output tokens, Qwen3-14B is the value inflection point in the Qwen3 dense line — cheaper than Qwen3-32B ($0.20 / $0.80) while stepping meaningfully above Qwen3-8B in quality on harder reasoning tasks.
| Type | Rate |
|---|---|
| Input | $0.378 / 1M tokens |
| Output | $1.51 / 1M tokens |
| Total parameters | 14.8B (13.2B non-embedding)huggingface.co/Qwen ↗ | verified |
| Architecture | Dense causal transformer, 40 layers, GQA 40/8 (Q/KV heads)huggingface.co/Qwen ↗ | verified |
| Context length | 131,072 tokens (32,768 native; extended to 131,072 via YaRN)huggingface.co/Qwen ↗ | verified |
| Max completion tokens | 32,768huggingface.co/Qwen ↗ | verified |
| License | Apache 2.0huggingface.co/Qwen ↗ | verified |
| Training stage | Pretraining & Post-training (instruct model)huggingface.co/Qwen ↗ | verified |
| Effective efficiency vs. prior gen | Matches Qwen2.5-32B-Base at 14B params — one full generation of parameter efficiency gainedqwenlm.github.io ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU Base model; 5-shot | 81.05% | arxiv.org — Qwen3 technical report ↗ | |
MMLU-Pro Base model | 61.03% | arxiv.org — Qwen3 technical report ↗ | |
BBH (BIG-Bench Hard) Base model; 3-shot CoT | 81.07% | arxiv.org — Qwen3 technical report ↗ | |
GSM8K Base model; 4-shot | 92.49% | arxiv.org — Qwen3 technical report ↗ | |
GPQA Base model; graduate-level science questions | 39.90% | arxiv.org — Qwen3 technical report ↗ |
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/qwen3-14b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Alibaba Qwen released the full Qwen3 suite including Qwen3-14B as a 14.8B dense causal LM with hybrid thinking modes, 128K context (YaRN-extended), 119-language support, and Apache 2.0 licensing. The announcement highlighted that Qwen3-14B base achieves parity with Qwen2.5-32B-Base, representing a full generation of parameter-efficiency improvement.
The Qwen Team released the full technical report (arXiv 2505.09388) covering pretraining, post-training, hybrid thinking architecture, multilingual capability, and comparative benchmarks across the Qwen3 model family including Qwen3-14B. The report documents base model benchmarks showing Qwen3-14B at MMLU 81.05, BBH 81.07, GSM8K 92.49, and GPQA 39.90.
Qwen3-14B is the mid-range dense model in the Qwen3 line. It outperforms Qwen3-8B (36 layers, 32/8 GQA) on reasoning-heavy and multilingual tasks, while Qwen3-32B (64 layers, 64/8 GQA, 32.8B params) provides a higher quality ceiling for the hardest problems. In cost: Qwen3-8B is cheaper for simple tasks; Qwen3-14B ($0.084/$0.252 per M tokens) is the inflection point delivering Qwen2.5-32B-class capability; Qwen3-32B ($0.20/$0.80 per M tokens) is worth the premium only when 14B quality is insufficient.
Through TheRouter, use the reasoning parameter on /v1/chat/completions. Directly, set enable_thinking=True or enable_thinking=False in the chat template when self-hosting, or prepend /think or /no_think in the system prompt or first user message. Thinking mode is the default. When thinking mode is active, responses include a <think>...</think> block that must be parsed or stripped before displaying to end users.
The model's native context is 32,768 tokens, extended to 131,072 tokens via YaRN positional interpolation. TheRouter sets context_length to 131,072 and max_completion to 32,768. Performance at the far end of the YaRN range may degrade on very complex prompts; validate your largest inputs before production rollout.
Yes. Qwen3-14B supports tools and tool_choice for function calling as well as response_format for structured output, available through the standard OpenAI-compatible parameters on TheRouter. It maintains tool-calling capability in both thinking and non-thinking modes, making it suitable for agentic pipelines that benefit from optional reasoning traces alongside tool execution.
Yes. TheRouter exposes qwen/qwen3-14b through the OpenAI-compatible /v1/chat/completions endpoint. Set your client's base URL to the TheRouter endpoint, configure an API key, and set the model to qwen/qwen3-14b. Streaming, tools, and response_format are all supported.
This claim comes directly from the Qwen3 official announcement: 'Qwen3-1.7B/4B/8B/14B/32B-Base performs as well as Qwen2.5-3B/7B/14B/32B/72B-Base, respectively.' In other words, each Qwen3 dense size-tier matches the next-larger Qwen2.5 tier, so Qwen3-14B base equals Qwen2.5-32B base. This reflects advances in model architecture, training data volume, and training methodology across the Qwen3 generation rather than any single breakthrough.
| source | URL | retrieved | |
|---|---|---|---|
| Total parameters | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| Architecture | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| Context length | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| Max completion tokens | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| License | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| Training stage | huggingface.co/Qwen ↗ | 2026-06-09 | verified |
| Effective efficiency vs. prior gen | qwenlm.github.io ↗ | 2026-06-09 | verified |
| MMLU | arxiv.org — Qwen3 technical report ↗ | 2026-06-09 | verified |
| MMLU-Pro | arxiv.org — Qwen3 technical report ↗ | 2026-06-09 | verified |
| BBH (BIG-Bench Hard) | arxiv.org — Qwen3 technical report ↗ | 2026-06-09 | verified |
| GSM8K | arxiv.org — Qwen3 technical report ↗ | 2026-06-09 | verified |
| GPQA | arxiv.org — Qwen3 technical report ↗ | 2026-06-09 | verified |
| Qwen3 family released — Qwen3-14B among six open-weight dense models | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Qwen3 technical report published on arXiv | arxiv.org ↗ | 2026-06-09 | verified |
| How does Qwen3-14B compare to Qwen3-32B and Qwen3-8B? | qwenlm.github.io ↗ | 2026-06-09 | to verify |
| How do I switch between thinking and non-thinking mode? | huggingface.co ↗ | 2026-06-09 | to verify |
| What is the effective context window for qwen/qwen3-14b on TheRouter? | huggingface.co ↗ | 2026-06-09 | to verify |
| Does Qwen3-14B support tool calling and structured output? | huggingface.co ↗ | 2026-06-09 | to verify |
| Can I use Qwen3-14B from Cursor, Cline, or another OpenAI-compatible client? | qwenlm.github.io ↗ | 2026-06-09 | to verify |
| Why is Qwen3-14B described as having Qwen2.5-32B-level capability? | qwenlm.github.io ↗ | 2026-06-09 | to verify |