You can deploy this model and start earning money today!

Thireus/
Kimi-K2.5-THIREUS-Q6_0-SPECIAL_SPLIT

Name: Kimi-K2.5-THIREUS-Q6_0-SPECIAL_SPLIT
Availability: InStock

Views

No views yet

Kimi-K2.5

🤔 What is this HuggingFace repository about?

This repository provides GGUF-quantized tensors for the Kimi-K2.5 model (official repo: https://huggingface.co/moonshotai/Kimi-K2.5). These GGUF shards are designed to be used with Thireus’ GGUF Tool Suite (https://gguf.thireus.com), a collection of tools that automatically finds the perplexity-optimal mix of quantizations for any given VRAM and RAM target. With this GGUF Tool Suite, you can produce your own Dynamic 3.0 Quants recipes and achieve optimum accuracy & SOTA quantization performance.

📖 Read more: https://github.com/Thireus/GGUF-Tool-Suite
🔍 Example of GGUF recipes: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples
🍳 Cook your own recipe files: https://gguf.thireus.com/quant_assign.html
☁️ Download GGUF models from recipe files: https://gguf.thireus.com/quant_downloader.html
📂 Browse available models: https://gguf.thireus.com

tl;dr: Expand the details section below

cd ~

# Make sure to install all ik_llama.cpp compilation dependencies...
apt install python3-dev python3-pip python3-venv python3-wheel python3-setuptools git acl netcat-openbsd cmake # pipx

# Obtain ik_llama's Thireus version - Windows/macOS/Linux builds available at https://github.com/Thireus/ik_llama.cpp/releases
git clone https://github.com/Thireus/ik_llama.cpp
cd ik_llama.cpp
git pull
# Build ik_llama.cpp
cmake -B build -DGGML_AVX=ON -DGGML_AVX2=ON -DLLAMA_CURL=OFF -DGGML_MAX_CONTEXTS=2048
cmake --build build --config Release -j16
cd ..

# Obtain Thireus' GGUF-Tool-Suite
GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/Thireus/GGUF-Tool-Suite

# Download model quant mix from recipe file - you can also try the web version: https://gguf.thireus.com/quant_downloader.html
cd GGUF-Tool-Suite
rm -f download.conf # Make sure to copy the relevant download.conf for the model before running quant_assign.py
cp -f models/Kimi-K2.5/download.conf . # Use the download.conf of the chosen model
mkdir -p kitchen && cd kitchen
../quant_downloader.sh ../recipe_examples/ik_harmonized_recipes/Kimi-K2.5.ROOT-1.9968bpw-3.8128ppl.238GB-GGUF_10GB-GPU_228GB-CPU.90e3c2f_4766d51.recipe

# Other recipe examples can be found at https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples

# Launch ik_llama's llama-cli:
ulimit -n 9999 # Lifts "too many open files" limitation on Linux
~/ik_llama.cpp/build/bin/llama-cli \
  -m Kimi-K2.5-THIREUS-BF16-SPECIAL_TENSOR-00001-of-01148.gguf \
  -mla 3 -fa auto -amb 512 -ctk f16 -c 4096 -ngl 99 \
  -ot "blk\.(3|4|5|6)\.ffn_.*=CUDA0" \
  -ot "blk\.(7|8|9|10)\.ffn_.*=CUDA1" \
  -ot exps=CPU -b 2048 -ub 1024 --warmup-batch --no-mmap --threads 36 \
  --main-gpu 0 \
  -p '<｜begin▁of▁sentence｜><｜User｜>What is the solution of x+5=-2?<｜Assistant｜><think>\n'

❓ Why does this Tool Suite exist?

Compatibility & Speed – unsloth’s dynamic quants may not always work optimally with ik_llama.cpp.
Custom Rig Fit – No off-the-shelf GGUF model perfectly matched my VRAM/RAM setup, so I built a way to tailor models and leverage extra VRAM/RAM to reduce perplexity.
Automated PPL-Optimal Quantization – To my knowledge, there was no open source flexible, automated method to minimize perplexity for any bits-per-weight (bpw) target—so I created one with excellent results!

📊 How does it compare to other GGUFs?

Here’s how DeepSeek-R1-0528 quantized with Thireus’ GGUF Tool Suite stacks up against other quantizers (lower perplexity = better at equal or lower bpw):

Note: The recipe_examples files illustrate good recipes. The Tool Suite computes the optimal ppl/bpw curve for you — just specify your target RAM, VRAM, and quant types, and quant_assign.py finds the best mix.

More perplexity/bpw graphs for other supported models: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/ppl_graphs

All PPL benchmarks are computed with the parameters -ctk f16 -c 512 -b 4096 -ub 4096. Changing any of these parameters will alter the PPL. In particular, reducing -b 4096 -ub 4096 increases the PPL, while increasing them decreases the PPL.

🚀 How do I get started?

Check out the GGUF Tool Suite README — focus on these sections:

⚠️ Requirements – Which ik_llama.cpp (or llama.cpp) version to use and how to compile.
- Windows binaries (no patching needed) at: https://github.com/Thireus/ik_llama.cpp/releases
📥 Download Model Shards – Use quant_downloader.sh or quant_downloader.html to fetch GGUF shards from any recipe.
- Recipe examples: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples
🧠 Run a Downloaded Model – Sample usage with llama-cli.
🛠️ Generate a Custom Recipe – Produce recipes tailored to your VRAM/RAM target usage for optimum perplexity.

✅ Supported Models

Supported models are listed under models/ in the Tool Suite Github repo. Presence of ppl_results.csv indicates official support and compatibility with quant_assign.py.

🤷‍♂️ Will I release baked dynamic quant GGUFs?

No, because I believe in tailored quantization for each user’s hardware. If you prefer ready-made shards, you are welcome to merge them via llama-gguf-split --merge, or request someone to publish them, or rely on generic GGUF dynamic quants such as unsloth's.

Instead, I prefer to share examples of recipes so users can see exactly how they were produced (command included inside these recipe files) and tweak them for their own rigs. The quant_downloader.sh script or quant_downloader.html (web port of this script) handles automatic fetching and verification of each shard. Note that recipes provided by Ubergarm on his model cards are also compatible with quant_downloader.sh and quant_downloader.html, providing a "SPECIAL_SPLIT" version of these models exists (see https://gguf.thireus.com/).

Users who don’t trust the GGUF shards on HuggingFace can also quantize their own by passing recipe lines to llama-quantize --custom-q (see example). Run llama-quantize --help to list compatible quants for quant_assign.py. This approach is especially useful if you prefer llama.cpp over ik_llama.cpp.

📦 What’s in this repository?

00001 GGUF header shard – Contains metadata (tokens, chat template, tensor count, etc.). This metadata can be explored directly from the HuggingFace web interface after clicking on that shard.
Tensor shards – Each shard holds one tensor; see tensors.map for names, quant types, sizes, SHA-256 hash, shard IDs, etc.
GPG-signed files – tensors.map and header shard are signed with the key in trusted-keys.asc for tamper detection.
Security note – Some papers about various ways to attack GGUFs and LLMs are available online, such as https://arxiv.org/abs/2505.23786, and there are also more classic security exploits like CVE-2024-23496 and CVE-2024-25664 through CVE-2024-25668. Only use GGUFs from reputable, trusted authors—or alternatively self-quantize—to avoid potential exploits.

💡 Pro Tips

You can easily download the BF16 model version to quantize your own shards:

mkdir kitchen  
echo '.*=bf16' > kitchen/bf16.recipe  
cd kitchen
../quant_downloader.sh bf16.recipe --qtype BF16

You can also quantize individual BF16 tensors without the need to download every BF16 .gguf shard:

BF16 model shards can also be individually quantized using a special version of ik_llama.cpp's llama-quantize utility which comes with the --individual-tensors option.

Source code: https://github.com/Thireus/ik_llama.cpp/tree/th/quantize_individual_tensors
Builds (macOS, Windows and Linux): https://github.com/Thireus/ik_llama.cpp/releases/tag/th-quantize_individual_tensors-b4210-7a44805

Usage example:

./llama-quantize --keep-split --imatrix imatrix_ubergarm.dat --individual-tensors 2,3,1094 Kimi-K2.5-THIREUS-BF16-SPECIAL_TENSOR-00001-of-01097.gguf my_new_shards.gguf iq3_s 12

For more information about how to use it: https://github.com/Thireus/GGUF-Tool-Suite/issues/45

You can produce your own quantized shards from Thireus' special BF16 model using quantize_model.sh found on https://github.com/Thireus/GGUF-Tool-Suite, for example:

./quantize_model.sh --model "Qwen3.5-122B-A10B" --qtype iq2_xxs

You can disable reasoning (thinking) when using jinja templates for supported models:

llama-server ... --jinja --chat-template-kwargs '{"enable_thinking": false}'

Enjoy optimized quantization! 🎉

Thireus/Kimi-K2.5-THIREUS-Q6_0-SPECIAL_SPLIT