mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_SPLIT

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+*.gguf.zbst filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,158 @@
 ---
 license: mit
 ---
+# mtp-Qwen3.5-397B-A17B
+
+## 🤔 What is this [HuggingFace repository](https://huggingface.co/Thireus/mtp-Qwen3.5-397B-A17B-THIREUS-BF16-SPECIAL_SPLIT/) about?
+
+This repository provides **GGUF-quantized tensors** for the [mtp](https://github.com/ggml-org/llama.cpp/blob/master/docs/speculative.md) layer(s) of the Qwen3.5-397B-A17B model (official repo: https://huggingface.co/Qwen/Qwen3.5-397B-A17B). These GGUF shards are designed to be used with **Thireus’ GGUF Tool Suite** (https://github.com/Thireus/GGUF-Tool-Suite), a collection of tools that automatically finds the perplexity-optimal mix of quantizations for any given a model size target. With this GGUF Tool Suite, you can produce your own Dynamic 3.0 Quants recipes and achieve optimum accuracy & SOTA quantization performance. Give it a try here: https://gguf.thireus.com/quant_assign.html  
+
+- 📖 Documentation: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/docs
+- 🔍 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://huggingface.co/Thireus/collections and https://gguf.thireus.com  
+
+*tl;dr: Expand the details section below*
+<details>
+
+```
+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/Qwen3.5-397B-A17B/download.conf . # Use the download.conf of the chosen model
+mkdir -p kitchen && cd kitchen
+# Obtain a recipe example for the chosen model from ../recipe_examples/
+../quant_downloader.sh ../recipe_examples/ik_llama.cpp_recipes/Qwen3.5-397B-A17B.ROOT-3.5993bpw-11.3565ppl.1GB-GGUF_0GB-GPU_0GB-CPU.9888e4b_831ff04.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-server \
+  -m Qwen3.5-397B-A17B-THIREUS-BF16-SPECIAL_TENSOR-00001-of-*.gguf \
+  -md mtp-Qwen3.5-397B-A17B-THIREUS-BF16-SPECIAL_TENSOR-00001-of-*.gguf --spec-type draft-mtp \
+  -fa auto -amb 1024 -ctk q8_0 -c 32768 -ngl 99 \
+  -b 4096 -ub 4096 --warmup-batch --no-mmap --threads 1 \
+  --main-gpu 0
+```
+
+</details>
+
+---
+
+## ❓ Why does this Tool Suite exist?
+
+1. **Compatibility & Speed** – [unsloth](https://huggingface.co/unsloth)’s dynamic quants may not always work optimally with `ik_llama.cpp`.  
+2. **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.  
+3. **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 Qwen3.5-397B-A17B quantized with **Thireus’ GGUF Tool Suite** stacks up against other quantizers (lower perplexity = better at equal or lower bpw):
+
+![PPLs Compared With Others](https://github.com/Thireus/GGUF-Tool-Suite/raw/main/ppl_graphs/Qwen3.5-397B-A17B.svg)
+
+> _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  
+
+*Qwen3.5 Thireus' PPL benchmarks are computed with the parameters `-ctk f16 -c 512 -b 512 -ub 512`. Changing any of these parameters will alter the PPL. In particular, reducing `-b 512 -ub 512` increases the PPL, while increasing them decreases the PPL.*
+
+---
+
+## 🚀 How do I get started?
+
+Check out the [GGUF Tool Suite README](https://github.com/Thireus/GGUF-Tool-Suite) — focus on these sections:
+
+1. ⚠️ **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  
+2. 📥 **Download Model Shards** – Use `quant_downloader.sh` or [quant_downloader.html](https://gguf.thireus.com/quant_downloader.html) to fetch GGUF shards from any recipe.  
+   - Recipe examples: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples  
+3. 🧠 **Run a Downloaded Model** – Sample usage with `llama-cli`.  
+4. 🛠️ **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](https://github.com/Thireus/GGUF-Tool-Suite/tree/main/models). 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](https://huggingface.co/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](https://gguf.thireus.com/quant_downloader.html) (web port of this script) handles automatic fetching and verification of each shard. Note that recipes provided by [Ubergarm](https://huggingface.co/ubergarm) on his model cards are also compatible with `quant_downloader.sh` and [quant_downloader.html](https://gguf.thireus.com/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](https://github.com/Thireus/GGUF-Tool-Suite/blob/main/models/DeepSeek-R1-0528/DeepSeek-R1-0528-THIREUS-ANY-SPECIAL.sh#L482-L486)). 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](https://github.com/Thireus/GGUF-Tool-Suite/blob/main/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-Thinking-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! 🎉

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00001-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9471bb585584b8106dee2e25a35aca304d410e663e04259698223eecbfa5324c
+size 10942976

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00002-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:509e71e0d53a34ca983cf4a354348887bf36b7d73bc269cbd265a7b42facd830
+size 271165600

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00003-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e604971d89b61fb79be3267c29cad382db6845e98f968bc1f42f5043736ad917
+size 271165632

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00004-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c2d96fe44cdc41d5faac6ddf0b2ecdcab4ac6fdf7f932e143bc2d475df5cd5c
+size 17891520

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00005-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76f7c03f097e16683be1921a27b2daf9d1fcb9d7242c854fa6975416e92c69c0
+size 8929472

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00006-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cac894a1cc792cf019f25843ed3f67dd83769ca3f229a1e74b59cbf857a9f6e6
+size 8929472

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00007-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2439e05c75a111d0cf04c507204a6de067cbb82098454d152c7fb9bd73241df0
+size 16544

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00008-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a383841b9c1a14f4526917d4ed05710344fc5583e4c826e0ad427618293c9eb
+size 16576

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00009-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ef6041a64bebeb84f632c7e38b6e1890f77613dbd4ec92966e6621068fd725a
+size 578814144

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00010-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d33d4468e6aec50eeb0af99762a3733b614500300b9b486daba97824931408d8
+size 572522688

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00011-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e246a7d01c74060a8c4bcec0e2312eb3d15936d3182fd00383a8c0ca1d571f1
+size 572522688

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00012-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5700fd87103c9525804cdcc87f8d25f5c2153bf1ec5db300de3abe011afc4a37
+size 4194496

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00013-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9107343270079620daa54d00f41607638f885acefdcf7c92865799ddb45a4de0
+size 1130688

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00014-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27fd94288aeb1c02e25dcaabe2ef3ebd0746756f17637a95ba7feb29656e3c16
+size 1118400

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00015-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a85a9a58663c0aeadc758232133e3a3626df07101c98a4df0208a31a25f5ebe
+size 1118400

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00016-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da01c2854a2b2689da9016560432916f40907d53b4383facc193b030690d117b
+size 8384

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00017-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31c12485ad7302732e14c61468643b9646c5101ef86ce09416a040dbcc73ab28
+size 16576

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00018-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1dd08c4fa66cb1ef4b884ca61d1f39ae308a21aa18f5bfb5cdb3e4ebbd3ed30
+size 1216

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00019-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf942018b7281b7b70863b43a73874ccbc4b1621a29eb6a3585270c288ed7303
+size 559296

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00020-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13d11323507a6ae7fcc950d01198c78d26d71b589c25c7489ee75d2679023419
+size 1216

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00021-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cbd7847e76ee5779b6b65da02f5e79c7dfe2298f5ca8f6455b75078f72fd6ae
+size 559296

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00022-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f9192fcd24af94df159de8e20ac22c4c26e37558e27b7adc86322abae79ed08
+size 16576

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00023-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64ef3cc4d66ab84b84ec15a3a3f37257914b73ab1446bb4d85484bc7d089ba13
+size 16576

mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00024-of-00024.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3770f6f1db4ce7763dcf6745cfe274c59d7e4d4cdc1031ce6c186e015dd95da7
+size 16576

tensors.map

@@ -0,0 +1,23 @@
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00002-of-00024.gguf:509e71e0d53a34ca983cf4a354348887bf36b7d73bc269cbd265a7b42facd830:output.weight:shape=(4096, 248320):dtype=iq2_kt:elements=1017118720:bytes=271165440
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00003-of-00024.gguf:e604971d89b61fb79be3267c29cad382db6845e98f968bc1f42f5043736ad917:token_embd.weight:shape=(4096, 248320):dtype=iq2_kt:elements=1017118720:bytes=271165440
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00004-of-00024.gguf:0c2d96fe44cdc41d5faac6ddf0b2ecdcab4ac6fdf7f932e143bc2d475df5cd5c:blk.60.attn_q.weight:shape=(4096, 16384):dtype=iq2_kt:elements=67108864:bytes=17891328
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00005-of-00024.gguf:76f7c03f097e16683be1921a27b2daf9d1fcb9d7242c854fa6975416e92c69c0:blk.60.nextn.eh_proj.weight:shape=(8192, 4096):dtype=iq2_kt:elements=33554432:bytes=8929280
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00006-of-00024.gguf:cac894a1cc792cf019f25843ed3f67dd83769ca3f229a1e74b59cbf857a9f6e6:blk.60.attn_output.weight:shape=(8192, 4096):dtype=iq2_kt:elements=33554432:bytes=8929280
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00007-of-00024.gguf:2439e05c75a111d0cf04c507204a6de067cbb82098454d152c7fb9bd73241df0:output_norm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00008-of-00024.gguf:1a383841b9c1a14f4526917d4ed05710344fc5583e4c826e0ad427618293c9eb:blk.60.attn_norm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00009-of-00024.gguf:3ef6041a64bebeb84f632c7e38b6e1890f77613dbd4ec92966e6621068fd725a:blk.60.ffn_down_exps.weight:shape=(1024, 4096, 512):dtype=iq2_kt:elements=2147483648:bytes=578813952
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00010-of-00024.gguf:d33d4468e6aec50eeb0af99762a3733b614500300b9b486daba97824931408d8:blk.60.ffn_gate_exps.weight:shape=(4096, 1024, 512):dtype=iq2_kt:elements=2147483648:bytes=572522496
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00011-of-00024.gguf:0e246a7d01c74060a8c4bcec0e2312eb3d15936d3182fd00383a8c0ca1d571f1:blk.60.ffn_up_exps.weight:shape=(4096, 1024, 512):dtype=iq2_kt:elements=2147483648:bytes=572522496
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00012-of-00024.gguf:5700fd87103c9525804cdcc87f8d25f5c2153bf1ec5db300de3abe011afc4a37:blk.60.ffn_gate_inp.weight:shape=(4096, 512):dtype=bf16:elements=2097152:bytes=4194304
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00013-of-00024.gguf:9107343270079620daa54d00f41607638f885acefdcf7c92865799ddb45a4de0:blk.60.ffn_down_shexp.weight:shape=(1024, 4096):dtype=iq2_kt:elements=4194304:bytes=1130496
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00014-of-00024.gguf:27fd94288aeb1c02e25dcaabe2ef3ebd0746756f17637a95ba7feb29656e3c16:blk.60.ffn_gate_shexp.weight:shape=(4096, 1024):dtype=iq2_kt:elements=4194304:bytes=1118208
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00015-of-00024.gguf:0a85a9a58663c0aeadc758232133e3a3626df07101c98a4df0208a31a25f5ebe:blk.60.ffn_up_shexp.weight:shape=(4096, 1024):dtype=iq2_kt:elements=4194304:bytes=1118208
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00016-of-00024.gguf:da01c2854a2b2689da9016560432916f40907d53b4383facc193b030690d117b:blk.60.ffn_gate_inp_shexp.weight:shape=(4096,):dtype=bf16:elements=4096:bytes=8192
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00017-of-00024.gguf:31c12485ad7302732e14c61468643b9646c5101ef86ce09416a040dbcc73ab28:blk.60.post_attention_norm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00018-of-00024.gguf:b1dd08c4fa66cb1ef4b884ca61d1f39ae308a21aa18f5bfb5cdb3e4ebbd3ed30:blk.60.attn_k_norm.weight:shape=(256,):dtype=f32:elements=256:bytes=1024
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00019-of-00024.gguf:cf942018b7281b7b70863b43a73874ccbc4b1621a29eb6a3585270c288ed7303:blk.60.attn_k.weight:shape=(4096, 512):dtype=iq2_kt:elements=2097152:bytes=559104
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00020-of-00024.gguf:13d11323507a6ae7fcc950d01198c78d26d71b589c25c7489ee75d2679023419:blk.60.attn_q_norm.weight:shape=(256,):dtype=f32:elements=256:bytes=1024
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00021-of-00024.gguf:6cbd7847e76ee5779b6b65da02f5e79c7dfe2298f5ca8f6455b75078f72fd6ae:blk.60.attn_v.weight:shape=(4096, 512):dtype=iq2_kt:elements=2097152:bytes=559104
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00022-of-00024.gguf:5f9192fcd24af94df159de8e20ac22c4c26e37558e27b7adc86322abae79ed08:blk.60.nextn.shared_head_norm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00023-of-00024.gguf:64ef3cc4d66ab84b84ec15a3a3f37257914b73ab1446bb4d85484bc7d089ba13:blk.60.nextn.enorm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384
+mtp-Qwen3.5-397B-A17B-THIREUS-IQ2_KT-SPECIAL_TENSOR-00024-of-00024.gguf:3770f6f1db4ce7763dcf6745cfe274c59d7e4d4cdc1031ce6c186e015dd95da7:blk.60.nextn.hnorm.weight:shape=(4096,):dtype=f32:elements=4096:bytes=16384