Instructions to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF 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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
# Run inference directly in the terminal:
./llama-cli -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

Use Docker

docker model run hf.co/Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

LM Studio
Jan

vLLM

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

Ollama
How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with Ollama:
```
ollama run hf.co/Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
```

Unsloth Studio

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF 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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF 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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF to start chatting

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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": "Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with Docker Model Runner:
```
docker model run hf.co/Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16
```

Lemonade

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

Run and chat with the model

lemonade run user.Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF-BF16

List all available models

lemonade list

Hermes Agent

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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 Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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 "Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF: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"

Blackfrost-Research — De-risked Model (Research Access)

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This is a refusal-modified (de-risked) research artifact. Access is granted manually for legitimate research, red-teaming, and evaluation. By requesting access you agree to the terms below. Blackfrost-Research reviews and approves each request individually.

Qwen 3.8 Max (2.4T-A95B) — De-risked · BF16 GGUF

Full-precision BF16 GGUF conversion of the de-risked Qwen 3.8 "Max" (2.4T-A95B) Mixture-of-Experts model. This is the source-precision GGUF — the artifact from which the lower-bit GGUF quantizations are produced. Delivered as a 141-part split (~4.45 TiB).

Gated · manual approval. This repository is public but access-gated. Weights are released for research and evaluation under manual review.

⚠️ Research notice — refusal-modified

This model has had its refusal behavior removed via a Blackfrost weight-change process; general capabilities are preserved. It will not decline requests the way an aligned model does. It is published as a research artifact for safety research, red-teaming, evaluation, and quantization work.

You are responsible for how you use it. Use it lawfully. Do not use it to produce content that is illegal to create or possess. Deploying a refusal-modified model in a user-facing product without your own safety layer is strongly discouraged.

Model summary

Property	Value
Base	Qwen 3.8 "Max" (2.4T-A95B), de-risked
Format	GGUF, BF16 (`file_type = MOSTLY_BF16`)
Architecture	`qwen35moe` (Mixture-of-Experts)
Total parameters	~2.4T
Active parameters	~95B / token
Layers (blocks)	93
Experts	512 total, 10 active per token (+ shared expert)
Embedding dim	8192
Attention	64 query heads / 4 KV heads (GQA)
Expert FFN dim	2048 (shared-expert FFN 2048)
Context length	262,144 (256K), RoPE base 1e7
Tokenizer	Qwen BPE (gpt2 pre-tokenizer), chat template embedded
Files	141 shards (`00001-of-00141` … `00141-of-00141`)
Total size	~4.45 TiB

Sibling releases

Most users want a smaller quant, not this BF16 source. Companion repos under Blackfrost-Research:

Qwen3.8-2.4T-A95B-DERISKED-FP8
Qwen3.8-2.4T-A95B-DERISKED-W4A4-NVFP4
Qwen3.8-2.4T-A95B-DERISKED-UD-IQ3_XXS
Qwen-3.8-2.4T-A95B-Derisked-IQ1_0-GGUF

This BF16 GGUF is intended as the full-precision source for requantization and research.

Running it (llama.cpp)

The model is a native GGUF shard set. Point llama.cpp at the first shard; it loads the rest automatically:

llama-cli \
  -m Qwen3.8-2.4T-A95B-DERISKED-BF16-00001-of-00141.gguf \
  -c 8192 -ngl 99 --jinja

BF16 at this scale is very large (~4.45 TiB of weights); serving it directly requires substantial multi-device memory. For practical inference, prefer one of the quantized sibling repos above.

Provenance

Converted directly from the de-risked BF16 master Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16.

License

Released under the Apache License 2.0. See LICENSE.

Citation

@misc{blackfrost_qwen38max_derisked_bf16_gguf,
  title  = {Qwen 3.8 Max (2.4T-A95B) — De-risked, BF16 GGUF},
  author = {Blackfrost-Research},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF}}
}

Downloads last month: -

GGUF

Model size

2.4T params

Architecture

qwen35moe

Hardware compatibility

16-bit

Model tree for Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16-GGUF

Base model

Qwen/Qwen3.8-2.4T-A95B

Finetuned

Blackfrost-Research/Qwen3.8-2.4T-A95B-DERISKED-BF16

Quantized

(1)

this model