AI Briefing
KO

MiMo-V2.5 GGUF Released

·2026.04.29 14:24

Key point

XiaomiMiMo's MiMo-V2.5 has been quantized into text-only GGUF and released.

1 / 2

Details

The 309B parameter mimo2 architecture model XiaomiMiMo/MiMo-V2.5 was quantized into text-only GGUF.

Image and audio inputs are not included, and multimodal support cannot be used until it is merged into llama.cpp upstream. For attention_value_scale support, building from the PR #22493 branch is recommended over the latest llama.cpp main.

Mixed quantization was applied focusing on the FFN tensors that account for a large portion of the MoE structure, designed to reduce overall size while suppressing quality degradation.

  • Q8_0: 305.68 GiB, PPL 5.135221, +0.1229% vs base
  • Q5_K_M: 212.42 GiB, PPL 5.142293, +0.2608%
  • Q4_K_M: 176.70 GiB, PPL 5.204791, +1.4793%
  • IQ4_XS: 136.78 GiB, PPL 5.270911, +2.7685%
  • IQ3_S: 105.33 GiB, PPL 5.547074, +8.1529%

KLD and PPL curves were also released together, allowing comparison of the size-quality tradeoffs of each quantization.

ggufllama-cppmimo-v2-5
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

HuggingFace BlogHugging Face Adds GGUF Model Support to TransformersRedditHugging Face Introduces Native GGUF Support in TransformersRedditDynamic Quantiser Released, Improving Accuracy Over Standard Quantization in VRAM-Constrained EnvironmentsNVIDIA BlogNVIDIA Unveils Local AI Acceleration and RTX Spark PCs at IFA 2026TLDR AIUnsloth Dynamic 3.0 GGUF ReleasedHacker NewsUnsloth Dynamic 3.0 GGUFs