Nanbeige4.2-3B-oQ6e

Nanbeige4.2-3B oQ6e

Apple Silicon Optimized oQ6e MLX Quantized Release

This repository contains an * oQ6e mixed-precision MLX quantized** version of Nanbeige4.2-3B, optimized for fast and memory-efficient local inference on Apple Silicon.

The original Nanbeige4.2-3B model was developed by the Nanbeige LLM Lab. This repository contains an optimized MLX oQ6e conversion only and does not include any additional fine-tuning or retraining.

Using * oQ6e sensitivity-aware mixed-precision quantization**, this release preserves the strong agentic reasoning, coding, and tool-use capabilities of the original model while significantly reducing memory usage for local deployment.

---

About Nanbeige4.2

Nanbeige4.2-3B is a compact agentic language model built upon Nanbeige4.2-3B-Base.

Despite having only 3B non-embedding parameters, it demonstrates remarkably strong performance across agentic reasoning, coding, office automation, and tool-use tasks.

A major innovation of Nanbeige4.2 is its Looped Transformer architecture, where hidden representations are passed back through the transformer stack after a complete forward pass. This architectural reuse increases effective model capacity without increasing parameter count.

The original model combines:

• Large-scale supervised fine-tuning
• Real-world environment interaction
• Synthetic environment generation
• Outcome-based reinforcement learning
• Process reward optimization
• Tool-use supervision
• Code-agent training
• Office-agent workflows

The result is a compact model capable of competing with significantly larger language models on many practical agent benchmarks.

---

Quantization

This release uses * oQ6e mixed-precision quantization**.

Specifications

• Format: MLX
• Quantization: oQ6e
• Method: Sensitivity-Aware Mixed Precision
• Target Platform: Apple Silicon
• Inference Engine: MLX / oMLX

Unlike conventional fixed-bit quantization, * oQ6e dynamically allocates precision according to layer sensitivity**, preserving critical reasoning and planning layers while aggressively compressing less sensitive portions of the network.

Benefits include:

• Excellent reasoning preservation
• Strong coding performance
• Efficient tool use
• Reduced memory usage
• Faster inference
• Optimized Apple Silicon execution

---

Example Usage

from mlx_lm import load, generate

model, tokenizer = load("yugeshkarunamurthy/Nanbeige4.2-3B oQ6e")

messages = [
    {
        "role": "user",
        "content": "Design a task planner for a coding assistant."
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    temp=0.7,
    top_p=0.95,
    top_k=20,
    max_tokens=16384,
)

print(response)

---

Optimized For

This release is optimized for:

• Apple M1
• Apple M2
• Apple M3
• Apple M4

Compatible with:

• MLX
• MLX-LM
• oMLX
• Open WebUI
• LM Studio (MLX)
• Local AI Agent Frameworks

---

Model Highlights

• Looped Transformer Architecture
• Strong Agentic Reasoning
• Native Tool Use
• Function Calling
• Software Engineering
• Office Automation
• Multi-step Planning
• Coding Assistance
• Research Assistance
• Instruction Following

---

Intended Use

Nanbeige4.2-3B oQ6e is well suited for:

• AI Agents
• Coding Assistants
• Tool Calling
• Software Development
• Research Automation
• Multi-step Planning
• Office Productivity
• Workflow Automation
• Local AI Applications
• General Purpose Reasoning

---

Hardware Recommendations

Recommended systems:

• Apple M1 Pro / Max / Ultra
• Apple M2 Pro / Max / Ultra
• Apple M3 Series
• Apple M4 Series

The oQ6e quantization significantly reduces memory usage, making the model practical for a wide range of Apple Silicon devices while maintaining excellent reasoning quality.

---

About oQ6e Quantization

oQ6e is a sensitivity-aware mixed-precision quantization technique designed to preserve model quality while substantially reducing memory requirements.

Rather than assigning identical precision to every layer, oQ6e analyzes layer sensitivity and selectively preserves higher precision where it has the greatest impact on reasoning and generation quality.

Benefits include:

• Better reasoning retention
• Stronger planning capabilities
• Higher coding accuracy
• Lower RAM usage
• Faster inference
• Excellent Apple Silicon optimization

---

Original Model

Nanbeige4.2-3B introduces a Looped Transformer architecture, where hidden states are reused through the transformer layers to increase effective model capacity without increasing parameter count. The model is trained with a combination of supervised fine-tuning and reinforcement learning to strengthen agentic behavior, tool use, and coding performance. It is released under the Apache-2.0 license. :contentReference[oaicite:0]{index=0}

---

Credits

Original Model

All credit for the original model, datasets, training methodology, evaluation, benchmarks, and research belongs entirely to:

Nanbeige LLM Lab

Original Repository:

https://huggingface.co/Nanbeige/Nanbeige4.2-3B

Base Model:

https://huggingface.co/Nanbeige/Nanbeige4.2-3B-Base

---

oQ6e MLX Quantized Release

This repository provides an Apple Silicon optimized * oQ6e MLX quantized** version of the original model.

No additional fine-tuning has been performed.

---

Acknowledgements

• Nanbeige LLM Lab
• Apple MLX
• Hugging Face
• Transformers
• vLLM
• SGLang
• MLX-LM
• oMLX
• OptiQ Quantization

---

Citation

If you use this model in research, please cite the original Nanbeige4.2 model:

@misc{nanbeige2026,
  title        = {Nanbeige4.2-3B},
  author       = {Nanbeige Team},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanbeige/Nanbeige4.2-3B}}
}

---

License

This release inherits the Apache-2.0 license from the original model.

Please refer to the original repository for complete licensing information.

---

Disclaimer

This repository contains an optimized * oQ6e MLX quantized conversion** intended for efficient local inference on Apple Silicon.

All original model architecture, datasets, training methodology, benchmarks, evaluations, and research remain entirely the work of the original Nanbeige LLM Lab.