Instructions to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT",
	filename="Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_TENSOR-00001-of-01132.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
# Run inference directly in the terminal:
llama-cli -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
# Run inference directly in the terminal:
llama-cli -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
# Run inference directly in the terminal:
./llama-cli -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Use Docker

docker model run hf.co/Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

LM Studio
Jan
Ollama
How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Ollama:
```
ollama run hf.co/Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
```

Unsloth Studio

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT to start chatting

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Run Hermes

hermes

Atomic Chat new
Docker Model Runner
How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Docker Model Runner:
```
docker model run hf.co/Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16
```

Lemonade

How to use Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT:BF16

Run and chat with the model

lemonade run user.Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT-BF16

List all available models

lemonade list

Thireus commited on Dec 30, 2025

Commit

0d05512

1 Parent(s): 19e3e1e

Update README

Browse files

Files changed (1) hide show

README.md +9 -8

README.md CHANGED Viewed

@@ -5,12 +5,13 @@ license: mit
 ## 🤔 What is this [HuggingFace repository](https://huggingface.co/Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT/) about?
-This repository provides **GGUF-quantized tensors** for the Qwen3-235B-A22B-Instruct-2507 model (official repo: https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507). 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 the Tool Suite, you can generate and download custom quantization “recipes” effortlessly.
 - 📖 Read more: https://github.com/Thireus/GGUF-Tool-Suite
-- 🔍 Example quant mixes: https://github.com/Thireus/GGUF-Tool-Suite/tree/main/recipe_examples
-- 🛠️ Create your own recipe: https://colab.research.google.com/github/Thireus/GGUF-Tool-Suite/blob/main/quant_recipe_pipeline.ipynb
-- 📂 Browse available quant shards: https://huggingface.co/Thireus/collections
 *tl;dr: Expand the details section below*
 <details>
@@ -33,7 +34,7 @@ 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:
 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/DeepSeek-R1-0528/download.conf . # Use the download.conf of the chosen model
@@ -46,7 +47,7 @@ mkdir -p kitchen && cd kitchen
 ulimit -n 9999 # Lifts "too many open files" limitation on Linux
 ~/ik_llama.cpp/build/bin/llama-cli \
   -m DeepSeek-R1-0528-THIREUS-BF16-SPECIAL_TENSOR-00001-of-01148.gguf \
-  -mla 3 -fa -amb 512 -fmoe -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 \
@@ -86,7 +87,7 @@ Check out the [GGUF Tool Suite README](https://github.com/Thireus/GGUF-Tool-Suit
 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` 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.
@@ -103,7 +104,7 @@ Supported models are listed under `models/` in the [Tool Suite Github repo](http
 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 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`.
 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 is this [HuggingFace repository](https://huggingface.co/Thireus/Qwen3-235B-A22B-Instruct-2507-THIREUS-BF16-SPECIAL_SPLIT/) about?
+This repository provides **GGUF-quantized tensors** for the Qwen3-235B-A22B-Instruct-2507 model (official repo: https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507). 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
+- ☁️ Download GGUF models from recipe files: https://gguf.thireus.com/quant_downloader.html
+- 🛠️ Create your own recipes: https://colab.research.google.com/github/Thireus/GGUF-Tool-Suite/blob/main/quant_recipe_pipeline.ipynb
+- 📂 Browse available models: https://gguf.thireus.com
 *tl;dr: Expand the details section below*
 <details>
 # 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/DeepSeek-R1-0528/download.conf . # Use the download.conf of the chosen model
 ulimit -n 9999 # Lifts "too many open files" limitation on Linux
 ~/ik_llama.cpp/build/bin/llama-cli \
   -m DeepSeek-R1-0528-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 \
 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.
 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`.