Back to directory

Kimi K3 GGUF Structural Validation Report

An independent structural validation report on two split GGUF releases of the Kimi K3 model, checking descriptor counts, identities, and shapes between quantization variants without downloading any tensor payload bytes.

Kimi K3 GGUF Structural Validation Report

This repository is a case-study report of an independent structural validation run against two split GGUF variants of Unsloth's Kimi-K3-GGUF release, UD-IQ1_M and UD-Q4_K_XL, pinned to a specific immutable artifact revision. Both variants are reported as "structurally validated with limitations," with 2,573 descriptors each, matching normalized identities and shapes between the two quantizations, and remote inspection that accepted zero tensor payload bytes.

The report is explicit about what it does not establish: artifact-specific provenance is unavailable, and payload equality and integrity, quantization fidelity, tokenizer parity, and runtime parity are not checked, so the findings say nothing about numerical correctness, model quality, benchmark performance, or the safety of the full Kimi release. The repository publishes the full report, an evidence manifest, a claim registry, a reproducibility manifest, and a public artifact index, generated by a separately maintained validation engine called OMIV.

The report and documentation are licensed CC BY 4.0, and the verification shell script is Apache 2.0.

More in Models

See all 405
  • mini-AGI

    github.com

    A continual learning language model that trains from scratch on a single 8 GB VRAM GPU, assembling its own architecture and adapting its size to the hardware and data available rather than shipping as a fixed pretrained model.

  • An open-source contrastive-embedding pipeline that maps technical CVE vulnerability descriptions to formal cybersecurity compliance controls in NIST SP 800-53 and CMMC, published with its model generation, fine-tuning and evaluation code.

  • peekaboolean

    github.com

    A small vision-language model, fine-tuned with LoRA on a 500-million-parameter backbone, that answers typed multiple-choice, ordinal-score and yes/no questions about an image with calibrated probabilities, in about 400 milliseconds on a laptop.