Intel/MiroThinker-v1.5-30B-gguf-q2ks-mixed-AutoRound

Model Details

This model is a mixed GGUF Q2_K_S format of miromind-ai/MiroThinker-v1.5-30B generated by the intel/auto-round algorithm based on the paper SignRoundV2: Toward Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs.

In this configuration, the Embedding layer and lm-head layer fallback to 8 bits, and non-expert layers fallback to 4 bits.

Please follow the license of the original model.

How To Use

Llama.cpp inference

./llama-cli -hf Intel/MiroThinker-v1.5-30B-gguf-q2ks-mixed-AutoRound

Generate the model

Please use auto-round 0.9.2 or >=0.95 as 0.93/0.94 have device bugs that causing the quantization process very slow

import torch
from auto_round import AutoRound
from auto_round.utils import llm_load_model

model_name = "miromind-ai/MiroThinker-v1.5-30B"

model, tokenizer=llm_load_model(model_name,trust_remote_code=False,device="cpu")
layer_config = {}
for n, m in model.named_modules():
    if isinstance(m,torch.nn.Embedding):
        layer_config[n] = {"bits": 8}
    if isinstance(m, torch.nn.Linear):
        if n=="lm_head":
            layer_config[n] = {"bits": 8}
            continue
        if "expert" in n and "shared_experts" not in n:
            layer_config[n] = {"bits": 2}
        elif n != "lm_head":
            layer_config[n] = {"bits": 4}
            print(n, 4)

ar = AutoRound(model, tokenizer=tokenizer, iters=0, scheme="gguf:q2_k_s", layer_config=layer_config)
ar.quantize_and_save(format="gguf:q2_k_s", output_dir="./MiroThinker-v1.5-30B-gguf-q2ks-mixed")

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

Here are a couple of useful links to learn more about Intel's AI software:

• Intel Neural Compressor link
• Intel AutoRound link

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2025signroundv2,
  title={SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs},
  author={Cheng, Wenhua and Zhang, Weiwei and Guo, Heng and Shen, Haihao},
  journal={arXiv preprint arXiv:2512.04746},
  year={2025}
}

arXiv | GitHub