AI4Burmese Padauk GGUF
Mission: Open, free, and accessible Burmese AI so Myanmar is not left behind in the age of AI.
AI4Burmese Padauk GGUF is the quantized GGUF release of Padauk, built for efficient local inference through Ollama, llama.cpp, and OpenAI-compatible local APIs.
Padauk is part of the AI4Burmese initiative: an open effort to make Burmese AI more accessible, practical, and buildable for everyone, including users, students, researchers, and developers.
- Model repository:
WYNN747/ai4burmese-padauk-gguf
- Source adapter repo:
WYNN747/ai4burmese-padauk
- Website: https://ai4burmese.com/
- Project page: https://waiyannyeinnaing.com/projects/padauk
Identity
- Brand: AI4Burmese
- Model: Padauk
- Format: GGUF
- Category: Burmese-first agentic small language model
- Lineage: Part of the AI4Burmese open Burmese AI ecosystem
Padauk is a Burmese-first, agentic-optimized small language model based on Gemma 4, tuned for Myanmar-context understanding, user-intent interpretation, and tool-connected execution in practical assistant workflows.
- Developed by: Dr. Wai Yan Nyein Naing
- Shared by: WYNN747
- Focus: Burmese-first agentic assistant, tool calling, local deployment
- Base model family: Gemma 4
- Languages: Burmese (my), English (en)
- Best for: Burmese local assistants, MCP workflows, edge deployment, and practical AI use
Model Summary
- Model name:
ai4burmese-padauk-gguf
- Base model family: Gemma 4
- Model type: Burmese-first agentic small language model
- Format: GGUF quantized release
- Primary focus: Burmese context understanding, local inference, tool-enabled assistant workflows
- Intended runtimes: Ollama, llama.cpp, OpenAI-compatible local APIs, agent frameworks, and edge deployment
Why AI4Burmese
Burmese remains underserved in modern AI. Many users, students, and builders still lack open, practical, and deployable AI systems designed for Burmese language use.
AI4Burmese exists to help close that gap by supporting open Burmese AI models and tools that are:
- Open for learning, contribution, and extension
- Free and accessible for broader public use
- Practical for real workflows, not only demos
- Buildable so developers can adapt and deploy them
- Burmese-first while remaining useful in bilingual settings
This GGUF release helps make that mission more usable in practice through local, self-hosted, and edge-friendly deployment.
Model Description
Padauk is designed for a simple goal: a Burmese assistant should understand Burmese intent well enough to drive tools, APIs, and action systems more reliably in real-world workflows.
It is tuned for:
- concise and useful Burmese interaction
- better Myanmar-context intent understanding
- structured instruction following and agent-style orchestration
- function and tool calling workflows
- efficient local deployment on modest hardware
This GGUF release enables more privacy-preserving, cost-efficient, and offline or self-hosted deployment.
Why Padauk
- Burmese-first behavior: optimized for Burmese prompts and Myanmar context
- Agentic workflow readiness: tuned for action planning, structured prompting, and orchestration
- Tool calling compatibility: designed for APIs, MCP servers, skills, and OpenAI-style tool interfaces
- Local and edge friendly: GGUF delivery for Ollama and lightweight runtimes
- Practical deployment focus: intended for usable assistant systems, not only benchmark demos
Validated Release Artifact
- Validated file:
padauk-gemma-q8_0.gguf
- SHA256:
2080409fb04855e695fa15e18b5618d0312f3d801b83518268bb063f9f62047a
- Checksum file:
SHA256SUMS.txt
- Validation path: converted from source checkpoint to F16 GGUF, load-tested in standalone
llama.cpp, quantized to Q8_0, then load-tested again
Gemma 4 Runtime Note
This GGUF uses the Gemma 4 shared-KV layout. If you see an error like missing tensor 'blk.24.attn_k.weight', that usually means the runtime is too old, not that the model export is incomplete.
- Minimum supported standalone
llama.cpp: b8751
- Recommended standalone
llama.cpp: b8833
- Current
llama-cpp-python wheels may still vendor an older llama.cpp revision and can fail on this shared-KV model family
For release validation, use standalone llama.cpp rather than relying only on older Python wheels.
Intended Use
- Burmese AI assistants and local productivity agents
- Tool-calling chat systems and MCP-connected assistants
- Skill-based architectures and OpenAI-compatible local serving
- Edge or on-device assistant prototypes
- Burmese-first AI experiments and local deployment workflows
Example use cases
- Burmese personal assistant with agentic tool access
- Burmese task runner connected to APIs
- Local knowledge plus tool orchestration agent
- MCP-connected Burmese desktop assistant
- Open and accessible Burmese AI applications for community use
Run with Ollama
1ollama pull wynn747/ai4burmese-padauk-gguf
2ollama run wynn747/ai4burmese-padauk-gguf:latest
3ollama list
Download from Hugging Face
1hf download WYNN747/ai4burmese-padauk-gguf \
2 --include "*.gguf" "Modelfile" "README.md" "SHA256SUMS.txt" \
3 --local-dir ./ai4burmese-padauk-gguf
Run as an OpenAI-Compatible Local API
1cd /path/to/ai4burmese-padauk
2./setup_api.sh
3./setup_api.sh run
Default endpoint: http://127.0.0.1:11434/v1
Agent Framework Compatibility
Works with OpenAI-compatible runtimes including LangChain, LangGraph, LiteLLM, CrewAI, MCP hosts, and custom skill registries.
For best tool-calling behavior:
- keep system instructions strict
- keep tool schemas stable
- keep JSON contracts predictable
- frame Burmese tasks clearly
- validate outputs before high-stakes use
FAQ
What is AI4Burmese Padauk GGUF?
A quantized GGUF release of Padauk, a Burmese-first agentic small language model based on Gemma 4 for local and self-hosted inference.
How do I run Padauk locally?
Pull wynn747/ai4burmese-padauk-gguf with Ollama, then run it. It can also be served through OpenAI-compatible local APIs depending on your runtime setup.
What is it best for?
Burmese assistant workflows that require intent understanding and tool-connected interaction through APIs, MCP servers, and skills.
How is it different from Burmese-GPT?
Burmese-GPT is a more foundation-style Burmese text generation direction. Padauk is an agentic-optimized assistant model focused on practical workflows and tool use.
How is it different from Burmese-Coder-4B?
Burmese-Coder-4B is specialized for Burmese programming and technical coding workflows. Padauk targets general assistant and agentic use cases.
Limitations
- Small model scale limits complex long-horizon reasoning compared to larger frontier models
- Tool execution reliability depends on system prompt quality, schema design, and runtime implementation
- MCP, skill, and API compatibility refers to workflow compatibility, not guaranteed correctness for every third-party stack
- Performance varies by quantization choice, backend, and device resources
- Outputs should be reviewed before use in high-stakes settings
Citation
If you use Padauk in research or product work, please cite the model page and source adapter page:
1@misc{padauk2026gguf,
2 title = {AI4Burmese Padauk GGUF: Burmese-First Agentic Small Language Model for Local Deployment},
3 author = {Wai Yan Nyein Naing},
4 year = {2026},
5 url = {https://huggingface.co/WYNN747/ai4burmese-padauk-gguf}
6}
License
Gemma license, subject to the base model terms and any repository-specific release notes.
Author
Dr. Wai Yan Nyein Naing
AI researcher and builder focused on Burmese AI, small language models, agentic systems, and practical deployment for low-resource language communities.
Links
Related Work
- AI4Burmese
- Padauk technical project page
- AI4Burmese Padauk source adapter
- Burmese-Coder-4B
- Burmese language open-source AI work
