Qwen3-14B — Model Documentation
Published by Nextbit256 S.L. | Last updated: May 2026 Compliance reference: AI Act Regulation (EU) 2024/1689, Art. 53.1.c and 53.1.d
Model Identification
| Field | Detail |
|---|---|
| Full model name | Qwen3-14B |
| Developed by | Qwen Team, Alibaba Cloud |
| Release date | April 2025 |
| Model card | huggingface.co/Qwen/Qwen3-14B |
| Technical report | arXiv:2505.09388 |
| Developer blog | qwenlm.github.io/blog/qwen3 |
Architecture and Parameters
| Field | Detail |
|---|---|
| Architecture | Dense (decoder-only transformer) |
| Total parameters | 14.8B |
| Active parameters per token | N/A — dense model (all parameters active) |
| Non-embedding parameters | 13.2B |
| Layers | 40 |
| Attention | GQA — 40 Q heads, 8 KV heads |
| Native context length | 32,768 tokens |
| Extended context (YaRN) | 131,072 tokens |
| Precision | BF16 |
License and Commercial Use
License: Apache License 2.0
Apache 2.0 is a permissive open-source license that allows free commercial use, distribution, and modification, provided that the original copyright notice and license text are retained and any modifications are documented.
Nextbit's verification: Nextbit has reviewed the license terms applicable to Qwen3-14B and confirmed that serving this model via API under a commercial inference service is permitted under Apache 2.0. No separate commercial agreement with Alibaba Cloud / Qwen Team is required for this use case.
Restrictions relevant to users: None specific to this model beyond standard Apache 2.0 terms. Users of Nextbit's API who integrate model outputs into their own products remain responsible for their own compliance with applicable law and any downstream licensing requirements.
Training Data Summary
The following is a summary of publicly available information provided by the Qwen Team regarding training data used for the Qwen3 model family, including Qwen3-14B. This is the same training data as the rest of the Qwen3 family (see also: Qwen3-30B-A3B documentation); only the model architecture (dense 14B vs. MoE) differs.
Pretraining data
The Qwen3 model family was pretrained on approximately 36 trillion tokens across 119 languages and dialects. This represents roughly double the training data used for the preceding Qwen2.5 generation (~18 trillion tokens).
Training data sources include:
- Web content: Text extracted from publicly available web pages, processed and filtered for quality.
- Document data: PDF and structured document content processed using Qwen2.5-VL for text extraction.
- Synthetic data: Mathematics and programming data generated using Qwen2.5-Math and Qwen2.5-Coder to strengthen reasoning and coding capabilities.
Pretraining stages
Training was conducted across three sequential stages:
| Stage | Tokens | Focus |
|---|---|---|
| Stage 1 | >30 trillion | General language skills, 4K context window |
| Stage 2 | ~5 trillion | Knowledge-intensive domains: STEM, coding, reasoning |
| Stage 3 | Not specified | Long-context extension to 32K tokens, high-quality long-context data |
Post-training
Post-training included supervised fine-tuning and reinforcement learning to enable the model's dual thinking/non-thinking mode operation, instruction following, and tool use capabilities.
Knowledge cutoff
The Qwen Team has not published an explicit knowledge cutoff date for Qwen3. Based on the model's release in April 2025, the training data cutoff is estimated to be no later than early 2025, though this has not been officially confirmed. Users should treat this as an approximation.
What is not publicly available
The Qwen Team has not published:
- Exact composition percentages of data sources
- Specific datasets or corpora used
- Data filtering or deduplication methodology in detail
- Training compute (FLOPs or GPU-hours)
The primary public reference for training data is the Qwen3 Technical Report (arXiv:2505.09388) and the Qwen3 blog post.
Languages Supported
Qwen3-14B supports 119 languages and dialects, including but not limited to: Arabic, Chinese (Simplified and Traditional), English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Thai, Turkish, Ukrainian, and Vietnamese.
Full language list: see the Qwen3 Technical Report.
Intended Uses
This model is designed for:
- Complex reasoning tasks (mathematics, logic, scientific reasoning)
- Code generation and analysis
- General-purpose instruction following and dialogue
- Multilingual tasks and translation
- Agentic workflows with tool and function calling
- Creative writing and role-playing
The model supports two operation modes:
- Thinking mode: Extended internal reasoning before generating a response, suitable for complex tasks
- Non-thinking mode: Direct response generation, suitable for conversational and latency-sensitive applications
Systemic Risk Assessment
| Field | Detail |
|---|---|
| Training FLOPs | Not published by the Qwen Team |
| Estimated FLOPs (dense model) | ~2.5 × 10²⁴ (estimate based on 14.8B parameters × ~36T tokens; formula: 6 × N × D) |
| Exceeds 10²⁵ FLOPs threshold? | No — estimated to be approximately 4× below the systemic risk threshold |
| Note on dense architecture | Qwen3-14B is a dense model; all 14.8B parameters are active per token. The full parameter count (not an active subset) is used for FLOPs estimation. |
| AI Office designation | Not designated as a systemic risk model as of May 2026 |
| Art. 55 obligations apply? | No |
This estimate is based on publicly available information. The Qwen Team has not published official training compute figures. If such figures are published, this section will be updated.
This documentation is published by Nextbit256 S.L. in accordance with Article 53(1)(c) and 53(1)(d) of Regulation (EU) 2024/1689 (AI Act). Nextbit256 S.L. serves this model via its inference API but did not develop or train it. All training data and architectural information is sourced from public documentation published by the Qwen Team / Alibaba Cloud.
For questions: [email protected]