| --- |
| language: |
| - en |
| - ar |
| library_name: transformers |
| license: apache-2.0 |
| pipeline_tag: text-generation |
| base_model: |
| - Qwen/Qwen3-235B-A22B-Instruct-2507 |
| tags: |
| - arabic |
| - saudi-dialect |
| - conversational |
| - qwen3 |
| - mixture-of-experts |
| - long-context |
| - text-generation |
| - ai-agents |
| - enterprise-ai |
| - cloud-ai |
| --- |
| |
| <p align="center"> |
| <img src="./vivo.png" alt="vivo-c-v1" width="100%"> |
| </p> |
|
|
| <div align="center"> |
|
|
| # vivo-c-v1 |
|
|
| ### Balanced Arabic Intelligence for Cloud and Enterprise Applications |
|
|
| [๐ VIVO AI](https://vivo.lahjai.net) โข |
| [๐ค Hugging Face](https://huggingface.co/wasmdashai/vivo-c-v1) โข |
| [๐ข Lahja AI](https://lahjai.net) |
|
|
| </div> |
|
|
| --- |
|
|
| ## Overview |
|
|
| **vivo-c-v1** is a large-scale conversational language model developed as part of the **VIVO AI** model family. |
|
|
| The model is built on **Qwen3-235B-A22B-Instruct-2507**, a Mixture-of-Experts causal language model with approximately **235 billion total parameters** and **22 billion activated parameters per token**. |
|
|
| vivo-c-v1 is being developed to provide a balanced foundation for Arabic conversational intelligence, enterprise assistants, AI agents, knowledge-based systems, cloud applications, and long-context workloads. |
|
|
| The model places particular emphasis on: |
|
|
| - Modern Standard Arabic |
| - Saudi and Gulf Arabic dialects |
| - Natural conversational interaction |
| - Long-context understanding |
| - Enterprise knowledge integration |
| - Tool and function calling |
| - Cloud-native deployment |
| - AI-agent workflows |
|
|
| --- |
|
|
| ## Why vivo-c-v1? |
|
|
| Many general-purpose language models are optimized primarily for broad multilingual benchmarks. **vivo-c-v1** is positioned around practical deployment scenarios where response quality, contextual continuity, scalability, and integration with external systems are essential. |
|
|
| The model is designed for applications that require: |
|
|
| - Natural Arabic conversations |
| - Regional dialect awareness |
| - Persistent conversational context |
| - Retrieval-Augmented Generation (RAG) |
| - Enterprise knowledge bases |
| - API and tool integration |
| - Intelligent workflow automation |
| - Scalable cloud inference |
|
|
| --- |
|
|
| ## Model Architecture |
|
|
| | Property | Value | |
| |---|---| |
| | Model type | Causal Language Model | |
| | Architecture | Qwen3 Mixture of Experts | |
| | Total parameters | Approximately 235B | |
| | Activated parameters | Approximately 22B per token | |
| | Non-embedding parameters | Approximately 234B | |
| | Number of layers | 94 | |
| | Attention heads | 64 query heads and 4 key-value heads | |
| | Number of experts | 128 | |
| | Activated experts | 8 | |
| | Native context length | 262,144 tokens | |
| | Extended context | Up to approximately 1,010,000 tokens | |
| | Tensor type | BF16 | |
| | License | Apache 2.0 | |
|
|
| > The parameter and architecture values above describe the underlying base architecture. They should not be interpreted as independently reproduced performance claims for vivo-c-v1. |
|
|
| --- |
|
|
| ## Core Capabilities |
|
|
| ### Arabic and Regional Dialects |
|
|
| vivo-c-v1 is intended to improve interactions for Arabic-speaking users by focusing on: |
|
|
| - Modern Standard Arabic |
| - Saudi Arabic dialects |
| - Gulf dialects |
| - Context-aware Arabic responses |
| - Reduced literal translation |
| - Better regional language adaptation |
| - Arabic instruction following |
|
|
| ### Long-Context Understanding |
|
|
| The underlying architecture supports a native context length of **262,144 tokens** and can be extended to approximately **1 million tokens** using supported long-context configurations. |
|
|
| This makes the model suitable for: |
|
|
| - Long enterprise documents |
| - Large knowledge bases |
| - Extended conversations |
| - Repository-level code analysis |
| - Research and technical documents |
| - Multi-step agent workflows |
|
|
| ### Enterprise and Agentic AI |
|
|
| vivo-c-v1 is designed for integration into: |
|
|
| - AI customer-service platforms |
| - Smart virtual assistants |
| - Enterprise search |
| - Knowledge management |
| - AI agents |
| - Tool-calling systems |
| - RAG pipelines |
| - Workflow automation |
| - Government, education, and healthcare applications |
|
|
| --- |
|
|
| ## Comparison with General Language Models |
|
|
| | Capability | vivo-c-v1 | General-purpose LLMs | |
| |---|---|---| |
| | Arabic-first deployment focus | High priority | General multilingual coverage | |
| | Saudi and Gulf dialect scenarios | Core target | Varies by model | |
| | Conversational applications | Primary use case | General text generation | |
| | Long-context workflows | Supported | Depends on model | |
| | Enterprise integration | APIs, RAG, tools, knowledge bases | Requires customization | |
| | AI-agent workflows | Designed for tool integration | Varies by implementation | |
| | Cloud deployment | Production-oriented | Depends on infrastructure | |
| | Resource balance | MoE with 22B active parameters | Architecture dependent | |
|
|
| --- |
|
|
| ## Quick Start |
|
|
| Install the latest compatible libraries: |
|
|
| ```bash |
| pip install --upgrade "transformers>=4.51.0" accelerate torch |
| ``` |
|
|
| Load the model with Hugging Face Transformers: |
|
|
| ```python |
| import torch |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_id = "wasmdashai/vivo-c-v1" |
| |
| tokenizer = AutoTokenizer.from_pretrained( |
| model_id, |
| trust_remote_code=True, |
| ) |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| torch_dtype="auto", |
| device_map="auto", |
| trust_remote_code=True, |
| ) |
| |
| messages = [ |
| { |
| "role": "system", |
| "content": ( |
| "You are an enterprise AI assistant specialized in Arabic " |
| "and Saudi conversational applications." |
| ), |
| }, |
| { |
| "role": "user", |
| "content": "ุตู
ู
ู
ุนู
ุงุฑูุฉ ูู
ูุตุฉ ู
ุณุงุนุฏ ุฐูู ู
ุคุณุณูุฉ ูุงุจูุฉ ููุชูุณุน.", |
| }, |
| ] |
| |
| text = tokenizer.apply_chat_template( |
| messages, |
| tokenize=False, |
| add_generation_prompt=True, |
| ) |
| |
| inputs = tokenizer( |
| [text], |
| return_tensors="pt", |
| ).to(model.device) |
| |
| with torch.inference_mode(): |
| generated_ids = model.generate( |
| **inputs, |
| max_new_tokens=2048, |
| temperature=0.7, |
| top_p=0.8, |
| top_k=20, |
| do_sample=True, |
| ) |
| |
| output_ids = generated_ids[0][inputs.input_ids.shape[-1]:] |
| |
| response = tokenizer.decode( |
| output_ids, |
| skip_special_tokens=True, |
| ) |
| |
| print(response) |
| ``` |
|
|
| --- |
|
|
| ## Cloud Deployment |
|
|
| ### vLLM |
|
|
| ```bash |
| vllm serve wasmdashai/vivo-c-v1 \ |
| --tensor-parallel-size 8 \ |
| --max-model-len 262144 |
| ``` |
|
|
| ### SGLang |
|
|
| ```bash |
| python -m sglang.launch_server \ |
| --model-path wasmdashai/vivo-c-v1 \ |
| --tp 8 \ |
| --context-length 262144 |
| ``` |
|
|
| For environments with limited GPU memory, reduce the context length: |
|
|
| ```bash |
| --max-model-len 32768 |
| ``` |
|
|
| or: |
|
|
| ```bash |
| --context-length 32768 |
| ``` |
|
|
| --- |
|
|
| ## Agentic Use |
|
|
| vivo-c-v1 can be integrated with: |
|
|
| - Function calling |
| - Model Context Protocol (MCP) |
| - Code execution tools |
| - Search and retrieval tools |
| - RAG systems |
| - Enterprise APIs |
| - Workflow orchestration |
| - Multi-agent applications |
|
|
| A typical production architecture may include: |
|
|
| ```text |
| User Interface |
| โ |
| โผ |
| API Gateway |
| โ |
| โผ |
| vivo-c-v1 Inference Service |
| โ |
| โโโ RAG and Vector Database |
| โโโ Enterprise Knowledge Bases |
| โโโ External APIs and Tools |
| โโโ Agent Orchestration |
| โโโ Monitoring and Safety Layer |
| ``` |
|
|
| --- |
|
|
| ## Long-Context Deployment |
|
|
| Processing contexts close to one million tokens requires substantial infrastructure for: |
|
|
| - Model weights |
| - KV cache |
| - Activation memory |
| - Tensor parallelism |
| - High-bandwidth GPU communication |
|
|
| For production deployment, context size should be selected according to the actual workload rather than always enabling the maximum supported length. |
|
|
| Recommended starting values: |
|
|
| | Workload | Suggested context | |
| |---|---:| |
| | Standard assistant | 16Kโ32K | |
| | Enterprise RAG | 32Kโ128K | |
| | Long-document analysis | 128Kโ256K | |
| | Specialized ultra-long context | Above 256K with dedicated infrastructure | |
|
|
| --- |
|
|
| ## Recommended Generation Settings |
|
|
| A practical starting configuration is: |
|
|
| ```python |
| generation_config = { |
| "temperature": 0.7, |
| "top_p": 0.8, |
| "top_k": 20, |
| "max_new_tokens": 2048, |
| "do_sample": True, |
| } |
| ``` |
|
|
| These values should be adjusted according to the application, latency target, output length, and factuality requirements. |
|
|
| --- |
|
|
| ## Intended Uses |
|
|
| vivo-c-v1 is intended for: |
|
|
| - Arabic conversational assistants |
| - Saudi and Gulf customer-service applications |
| - Enterprise copilots |
| - AI agents and automation |
| - Knowledge-base assistants |
| - RAG systems |
| - Long-document analysis |
| - Code and technical assistance |
| - Government digital services |
| - Education and healthcare platforms |
| - Cloud-native AI applications |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - The model may generate inaccurate or unsupported information. |
| - Arabic dialect quality may vary by topic and prompt. |
| - Long-context support does not guarantee perfect recall of every detail. |
| - Large-scale inference requires substantial GPU infrastructure. |
| - Generated code and business recommendations should be reviewed. |
| - Sensitive or high-stakes outputs require human validation. |
| - Performance results of the base model should not be presented as independently reproduced vivo-c-v1 results unless separate evaluations are published. |
|
|
| --- |
|
|
| ## Responsible Use |
|
|
| Users are responsible for: |
|
|
| - Reviewing generated content |
| - Protecting personal and confidential information |
| - Applying appropriate access controls |
| - Monitoring production outputs |
| - Testing integrations before deployment |
| - Complying with applicable laws and organizational policies |
|
|
| --- |
|
|
| ## Platform Links |
|
|
| ### vivo-c-v1 on Hugging Face |
|
|
| https://huggingface.co/wasmdashai/vivo-c-v1 |
|
|
| ### VIVO AI Platform |
|
|
| https://vivo.lahjai.net |
|
|
| ### Lahja AI |
|
|
| https://lahjai.net |
|
|
| --- |
|
|
| ## Base Model and Attribution |
|
|
| vivo-c-v1 is based on: |
|
|
| **Qwen3-235B-A22B-Instruct-2507** |
|
|
| https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507 |
|
|
| Please follow the applicable base-model license and attribution requirements. |
|
|
| --- |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{qwen3technicalreport, |
| title = {Qwen3 Technical Report}, |
| author = {Qwen Team}, |
| year = {2025}, |
| eprint = {2505.09388}, |
| archivePrefix= {arXiv}, |
| primaryClass = {cs.CL} |
| } |
| ``` |
|
|
| --- |
|
|
| ## License |
|
|
| This repository uses the **Apache License 2.0**, subject to the applicable terms of the base model and included dependencies. |
