Instructions to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
# Run inference directly in the terminal:
llama cli -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
# Run inference directly in the terminal:
llama cli -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP: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 random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
# Run inference directly in the terminal:
./llama-cli -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP: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 random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Use Docker

docker model run hf.co/random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

LM Studio
Jan
Ollama
How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with Ollama:
```
ollama run hf.co/random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
```

Unsloth Studio

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP 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 random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP 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 random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP to start chatting

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with Docker Model Runner:
```
docker model run hf.co/random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16
```

Lemonade

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Run and chat with the model

lemonade run user.Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP-BF16

List all available models

lemonade list

Hermes Agent

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP: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 random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "random257/Qwen3.5-122B-A10B-Wangzhang-Abliterix-No-MTP:BF16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

random257 commited on Jul 3

Commit

2bd86c4

verified ·

1 Parent(s): 29f463f

Update README.md

Browse files

Files changed (1) hide show

README.md +3 -3

README.md CHANGED Viewed

@@ -11,16 +11,16 @@ license_link: https://huggingface.co/Qwen/Qwen3.5-122B-A10B/blob/main/LICENSE
 > *   Best for testing; avoid public-facing applications.
 > **Note:**
-> *   The internal MTP layer was ditched via `--no-mtp`. The uploader was *too lazy* to patch the conversion bugs.
 > *   **Experimental conversions**; minor tensor misalignments, loading issues, or long-context quirks may occur. Use at your own risk.
 **BF16**: Directly converted from [wangzhang/Qwen3.5-122B-A10B-abliterix](https://huggingface.co/wangzhang/Qwen3.5-122B-A10B-abliterix) to GGUF (without MTP), split into multiple files, with updated `general.name`.
-**MQ68**: Mixed precision (Q6_K - Q8_0). Uses **Q6_K for large tensors** in expert blocks; BF16 for the output tensor; the rest remains at **Q8_0**. Inference speed is **0.8x** that of pure Q6_K.
 **MQ516_special**: Mixed precision (Q5_K - BF16). Uses **Q5_K** only for *large tensors* in *early SSM* blocks, **Q6_K** for large tensors in *other expert blocks*, and retains **BF16** for all *small-to-medium and critical* tensors (including the output projection). Since all medium-sized mandatory (dense) tensors are BF16, inference speed drops to **0.55x** that of pure Q6_K. *Experimental; the strategy is purely heuristic and unverified.*
-The upstream source model was altered by the Abliterix algorithm, causing degradation in emotional resonance, empathy, and literary creativity. It now exhibits a cold, mechanical personality (less severe than the 35B-A3B variant). This makes it highly effective for **critical review** tasks but suboptimal for creative writing. For models with restored emotional capabilities, please *use the Heretic variants* instead.
 Both MQ68 and MQ516_special can squeeze into a gfx1151-128G device. However on Windows, RAM offloading may cause further slowdowns compared to pure Q6_K (model size >96G, &tilde;107G at 262k context), while also increasing OOM risks that require careful management; unlike pure Q6_K (&tilde;104G at 262k context) which leaves sufficient margin and thus remains stable.

 > *   Best for testing; avoid public-facing applications.
 > **Note:**
+> *   The internal MTP layer was ditched via `--no-mtp`. The uploader was *too lazy* to patch the conversion bugs. May impact inference speed.
 > *   **Experimental conversions**; minor tensor misalignments, loading issues, or long-context quirks may occur. Use at your own risk.
 **BF16**: Directly converted from [wangzhang/Qwen3.5-122B-A10B-abliterix](https://huggingface.co/wangzhang/Qwen3.5-122B-A10B-abliterix) to GGUF (without MTP), split into multiple files, with updated `general.name`.
+**MQ68**: Mixed precision (Q6_K - Q8_0). Uses **Q6_K for large tensors** in expert blocks; **BF16** for the output tensor; the rest remains at **Q8_0**. Inference speed is **0.8x** that of pure Q6_K.
 **MQ516_special**: Mixed precision (Q5_K - BF16). Uses **Q5_K** only for *large tensors* in *early SSM* blocks, **Q6_K** for large tensors in *other expert blocks*, and retains **BF16** for all *small-to-medium and critical* tensors (including the output projection). Since all medium-sized mandatory (dense) tensors are BF16, inference speed drops to **0.55x** that of pure Q6_K. *Experimental; the strategy is purely heuristic and unverified.*
+The upstream source model was altered by the Abliterix algorithm, causing degradation in emotional resonance, empathy, and literary creativity. It now exhibits a slightly **cold**, **mechanical personality** (less severe than the 35B-A3B variant). This makes it effective for **critical review** tasks but suboptimal for creative writing. For models with restored emotional capabilities, please *use the Heretic variants* instead.
 Both MQ68 and MQ516_special can squeeze into a gfx1151-128G device. However on Windows, RAM offloading may cause further slowdowns compared to pure Q6_K (model size >96G, &tilde;107G at 262k context), while also increasing OOM risks that require careful management; unlike pure Q6_K (&tilde;104G at 262k context) which leaves sufficient margin and thus remains stable.