Aminlight/Qwen-AgentWorld-35B-A3B-4bit-MLX

Qwen-AgentWorld-35B-A3B Uniform 4-bit MLX

This repository contains an MLX-compatible uniform 4-bit quantization of Qwen/Qwen-AgentWorld-35B-A3B.

This is a quantized conversion of the original model, not a fine-tune.

Model Details

• Base model: Qwen/Qwen-AgentWorld-35B-A3B

• Base model type: Language World Model / causal language model

• Total parameters: 35B

• Activated parameters: 3B

• Context length: 262,144 tokens

• Quantization format: MLX

• Quantization method: uniform 4-bit

• Reference dtype during conversion: bfloat16

• Converter-reported bits per weight: 4.503

• Output model size: 18,605.3 MB

Quantization Summary

This model was produced as the uniform_4bit output during the MLX conversion process.

The relevant conversion output was:

Converting Qwen/Qwen-AgentWorld-35B-A3B → optiq_output/Qwen-AgentWorld-35B-A3B-OptiQ-6bpw-MLX/uniform_4bit
Quantization: uniform 4-bit
[INFO] Using dtype: bfloat16
[INFO] Quantizing
[INFO] Quantized model with 4.503 bits per weight.
Output model size: 18605.3 MB

Unlike the OptiQ mixed-precision version, this variant uses uniform 4-bit quantization rather than a per-layer mixed 4-bit / 8-bit allocation.

Usage

Install MLX LM:

pip install -U mlx-lm

Run generation:

python -m mlx_lm.generate \
  --model YOUR_USERNAME/Qwen-AgentWorld-35B-A3B-Uniform-4bit-MLX \
  --prompt "You are a language world model simulating a Linux terminal environment. Given the user's command, predict the terminal output.\n\nAction: execute_bash\nCommand: ls -la"

Python example:

from mlx_lm import load, generate

model, tokenizer = load("YOUR_USERNAME/Qwen-AgentWorld-35B-A3B-Uniform-4bit-MLX")

prompt = """You are a language world model simulating a Linux terminal environment.
Given the user's command, predict the terminal output.

Action: execute_bash
Command: ls -la
"""

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.6,
    verbose=True,
)

print(response)

Replace YOUR_USERNAME/Qwen-AgentWorld-35B-A3B-Uniform-4bit-MLX with the actual Hugging Face repository path.

Intended Use

This model is intended for local MLX inference and experimentation with agentic environment simulation. Example use cases include:

• local Apple Silicon inference

• agentic workflow prototyping

• terminal, tool-use, and environment-simulation experiments

• lightweight local testing compared with larger precision variants

• research and development with language world models

Limitations

This is a uniform 4-bit quantized version of the base model. It is expected to be smaller than the mixed-precision OptiQ version, but may show a larger quality drop compared with the original bfloat16 checkpoint or the OptiQ mixed-precision quantization.

Quantization may affect:

• factual accuracy

• reasoning consistency

• long-context behavior

• tool-use reliability

• environment-simulation fidelity

• code and terminal-output prediction quality

No independent benchmark evaluation of this quantized checkpoint is included. Users should evaluate it on their own tasks before relying on it in production or high-stakes settings.

Safety

This model inherits the safety profile and limitations of the original Qwen/Qwen-AgentWorld-35B-A3B model. Quantization does not add alignment, safety training, or new refusal behavior.

Do not use this model for high-stakes decisions without additional evaluation, monitoring, and human oversight.

License

This quantized model is released under the same license as the base model: Apache 2.0.

Please also consult the original model repository for authoritative license and usage details:

https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B

Acknowledgements

• Base model: Qwen/Qwen-AgentWorld-35B-A3B

• Original authors: Qwen

• Runtime format: MLX

• Quantization: uniform 4-bit