AI Transparency & Model Provider Notice

1. Integrator Role and Architecture Disclosure

In accordance with Vietnam's Law on Artificial Intelligence No. 134/2025/QH15 and Decree No. 142/2026/ND-CP, WinSell discloses that it operates as an AI System Deployer and intelligent API Gateway Router.

WinSell does NOT claim to own, build, or exclusively train the underlying large foundation models. These models are operated by world-leading independent third-party technology providers and accessed through authenticated API protocols.

2. Integrated Model Providers

WinSell provides access to 79 active foundational AI models across 17 world-leading technology organizations and developers, managed centrally and listed transparently in our public Model Pricing Catalog, including:

- OpenAI Inc. (USA): Next-generation reasoning and language models including GPT-6 (Astra), GPT-5.6 (Sol, Terra, Luna), GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5.3 Codex, GPT-5.2, GPT-OSS (120B, 20B), and high-resolution image models GPT-Image (2.5, 2, 1.5). Governed by OpenAI Business Terms.

- Anthropic PBC (USA): The complete Claude series including Claude Sonnet (Sonnet 5, 4.6, 4.5, 4), Claude Opus (Opus 5, 4.8, 4.7, 4.6), Claude Haiku 4.5, and Claude Fable (5.1, 5). Governed by Anthropic Commercial Terms.

- Google LLC / Alphabet (USA): Next-generation multimodal Gemini series (Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash, 3.5 Flash, Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, Gemini 3 Flash), Gemini Image models (3.1, 3, 2.5) with Nano Banana presets, and open-weights Gemma 4 (31B) via Google Cloud Vertex AI.

- xAI Corp. (USA): Grok series (Grok 4.6, Grok 4.5, Grok 4.3) and Grok Imagine Image generation models (Quality, Standard, Lite) governed by xAI Enterprise API Terms.

- DeepSeek AI (China): Next-generation DeepSeek V4 series (DeepSeek-V4.1 Flash, DeepSeek-V4 Pro, DeepSeek-V4 Flash) specialized for mathematics, coding, and logical reasoning.

- Zhipu AI (China): GLM series (GLM 5.3 Flash, GLM 5.3, GLM 5.2, GLM 5.1, GLM 5) supporting multimodal reasoning and native tool calls.

- Moonshot AI (China): Kimi K3 and Kimi K2.7 models engineered for ultra-long context document comprehension.

- Alibaba Cloud (China): Qwen series (Qwen 3.8 Flash, Qwen 3 Coder Next) optimized for commercial analytics and software development.

- MiniMax AI (China): MiniMax M3 and MiniMax M2.7 for conversational roleplay and summarization.

- Xiaomi Corporation (China): Mimo v2.5 Pro and Mimo v2.5 models optimized for device ecosystem efficiency.

- ByteDance / Volcano Engine (China): Seedream 5.0 Pro model specialized for high-fidelity image synthesis.

- Tencent Holdings (China): Hunyuan HY3 model for enterprise information retrieval and natural language dialogue.

- Meta Platforms Inc. (USA): Open-weights Llama 3.3 70B deployed on enterprise cloud infrastructure.

- NVIDIA Corporation (USA): Nemotron 3 Ultra enterprise models optimized for GPU-accelerated computing.

- Nomic AI Inc. (USA): Nomic Embed Text (v1.5, v1) and Nomic Embed Vision (v1.5, v1) for multimodal vector embeddings in RAG systems.

- Cursor / Anysphere Inc. (USA): Composer 2.5 specialized for integrated development environments and software engineering workflows.

- Muse AI (International): Muse Spark 1.3 and Muse Spark 1.2 models engineered for low-latency conversational responses.

Each model features distinct capabilities, latency, and CUBE consumption rates. Users select their preferred model inside Wisa or Koda.

3. Zero Training on Customer Data Policy

Under our enterprise API agreements with foundational model providers, prompts and customer content processed through WinSell are NOT used by upstream AI providers to train or fine-tune their public models. Specific rules concerning upstream 30-day abuse-monitoring retention and data deletion lifecycles are detailed in our Data Governance & Chat Retention Policy (/legal/data-governance).

4. AI Output Identification and Watermarking

In compliance with generative AI labeling rules, WinSell's user interfaces clearly denote active AI models and visually differentiate between human prompts and machine-generated responses.