Instructions to use FoolDev/Thanatos-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FoolDev/Thanatos-27B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="FoolDev/Thanatos-27B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("FoolDev/Thanatos-27B", dtype="auto")

llama-cpp-python

How to use FoolDev/Thanatos-27B with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="FoolDev/Thanatos-27B",
	filename="Thanatos-27B.Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": [
				{
					"type": "text",
					"text": "Describe this image in one sentence."
				},
				{
					"type": "image_url",
					"image_url": {
						"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
					}
				}
			]
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use FoolDev/Thanatos-27B with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf FoolDev/Thanatos-27B:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FoolDev/Thanatos-27B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf FoolDev/Thanatos-27B:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FoolDev/Thanatos-27B:Q4_K_M

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 FoolDev/Thanatos-27B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf FoolDev/Thanatos-27B:Q4_K_M

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 FoolDev/Thanatos-27B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf FoolDev/Thanatos-27B:Q4_K_M

Use Docker

docker model run hf.co/FoolDev/Thanatos-27B:Q4_K_M

LM Studio
Jan

vLLM

How to use FoolDev/Thanatos-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FoolDev/Thanatos-27B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FoolDev/Thanatos-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/FoolDev/Thanatos-27B:Q4_K_M

SGLang

How to use FoolDev/Thanatos-27B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "FoolDev/Thanatos-27B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FoolDev/Thanatos-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "FoolDev/Thanatos-27B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FoolDev/Thanatos-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Ollama
How to use FoolDev/Thanatos-27B with Ollama:
```
ollama run hf.co/FoolDev/Thanatos-27B:Q4_K_M
```

Unsloth Studio new

How to use FoolDev/Thanatos-27B 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 FoolDev/Thanatos-27B 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 FoolDev/Thanatos-27B to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for FoolDev/Thanatos-27B to start chatting

Pi new

How to use FoolDev/Thanatos-27B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf FoolDev/Thanatos-27B:Q4_K_M

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": "FoolDev/Thanatos-27B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use FoolDev/Thanatos-27B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf FoolDev/Thanatos-27B:Q4_K_M

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 FoolDev/Thanatos-27B:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use FoolDev/Thanatos-27B with Docker Model Runner:
```
docker model run hf.co/FoolDev/Thanatos-27B:Q4_K_M
```

Lemonade

How to use FoolDev/Thanatos-27B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull FoolDev/Thanatos-27B:Q4_K_M

Run and chat with the model

lemonade run user.Thanatos-27B-Q4_K_M

List all available models

lemonade list

FoolDev Claude Opus 4.7 commited on 10 days ago

Commit

732c3be

1 Parent(s): 693cf65

docs: correct MTP claim — not actually usable via llama.cpp / Ollama today

Browse files

Investigation triggered by an attempt to document the consumer recipe
for llama.cpp's new MTP speculative decoding (PR #22673, merged
2026-05-16). Three findings, all corroborating:

1. `convert_hf_to_gguf.py` explicitly skips MTP tensors for the
qwen35 / qwen35moe arch family — comment in source: "MTP
tensors are not used at inference yet; align with Qwen3Next
behaviour".
2. `src/models/qwen35.cpp` and `qwen35moe.cpp` contain zero MTP /
nextn references — even if tensors were preserved, the loader
wouldn't read them.
3. `gguf.GGUFReader` on both this repo's bundled GGUF and the
source unsloth/Qwen3.6-27B-GGUF Q4_K_M: 851 tensors each, no
mtp/draft/eagle/spec entries, final tensor is
blk.63.post_attention_norm.weight. The MTP head from the
upstream safetensors was dropped during conversion.

PR #22673's MTP support landed for other architectures, not qwen35.
The README's "Multi-token prediction (MTP) head trained for
speculative decoding" bullet, taken at face value, was misleading
users into thinking the speedup was available via llama.cpp.

Corrected bullet now distinguishes upstream-safetensors-via-vLLM/SGLang
(working) from llama.cpp / Ollama (not yet), points at the relevant
PR for tracking, and footnotes the empirical check. CHANGELOG picks
up a "Fixed" entry with the full evidence trail.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Files changed (2) hide show

CHANGELOG.md +28 -0
README.md +14 -1

CHANGELOG.md CHANGED Viewed

@@ -7,6 +7,34 @@ and documentation**, not the underlying base model.
 ## [Unreleased]
 ### Changed (3rd round trip — qwen35 → qwen36, user-directed despite audit)
 - **Bundle re-stamped `general.architecture: 'qwen35'` → `'qwen36'`**
   in `hf upload` commit `973d7ef` (HF). Third stamp flip on the

 ## [Unreleased]
+### Fixed
+- README "Multi-token prediction (MTP)" bullet corrected. The
+  earlier wording — "MTP head trained for speculative decoding" —
+  was technically true about the upstream `Qwen/Qwen3.6-27B`
+  safetensors but misleading for the GGUF bundle this repo ships
+  and for llama.cpp / Ollama users in general:
+  - **GGUFs are stripped.** `convert_hf_to_gguf.py` explicitly
+    skips MTP tensors for the `qwen35` / `qwen35moe` arch family
+    ("MTP tensors are not used at inference yet; align with
+    Qwen3Next behaviour"). Confirmed empirically via
+    `gguf.GGUFReader` on both `Thanatos-27B.Q4_K_M.qwen35.gguf`
+    and the source `unsloth/Qwen3.6-27B-GGUF` Q4_K_M: both have
+    851 tensors and zero entries matching `mtp.*` / `draft.*` /
+    `eagle.*` / `spec.*`. Last tensor in either is
+    `blk.63.post_attention_norm.weight` — the final layer norm,
+    no MTP head after it.
+  - **Loader doesn't support it.** `src/models/qwen35.cpp` and
+    `qwen35moe.cpp` contain no MTP / nextn references; even if
+    the tensors were in the GGUF, the loader wouldn't use them.
+  - **PR #22673's scope.** llama.cpp's MTP support (merged
+    2026-05-16) was added for other architectures, not the
+    `qwen35` family. The README bullet now says so explicitly,
+    points to vLLM (`qwen3_next_mtp`) and SGLang
+    (`--speculative-algo NEXTN`) as the working consumer
+    recipes against the safetensors, and notes that we're
+    tracking the follow-up that would extend MTP to qwen35 /
+    qwen35moe.
 ### Changed (3rd round trip — qwen35 → qwen36, user-directed despite audit)
 - **Bundle re-stamped `general.architecture: 'qwen35'` → `'qwen36'`**
   in `hf upload` commit `973d7ef` (HF). Third stamp flip on the

README.md CHANGED Viewed

@@ -182,7 +182,20 @@ If you want the safetensors for `transformers`, fetch them from [`Qwen/Qwen3.6-2
   `mmproj` projector (not redistributed here; pull `mmproj-F16.gguf`
   from `unsloth/Qwen3.6-27B-GGUF`). See [Vision](#vision) below for
   current loader compatibility.
-- Multi-token prediction (MTP) head trained for speculative decoding
 ### Stamp choice

   `mmproj` projector (not redistributed here; pull `mmproj-F16.gguf`
   from `unsloth/Qwen3.6-27B-GGUF`). See [Vision](#vision) below for
   current loader compatibility.
+- Multi-token prediction (MTP) head trained for speculative decoding —
+  present in the upstream `Qwen/Qwen3.6-27B` safetensors and usable via
+  vLLM (`qwen3_next_mtp`) or SGLang (`--speculative-algo NEXTN`).
+  **Not usable via llama.cpp / Ollama today**: the GGUF converter
+  (`convert_hf_to_gguf.py`) explicitly skips MTP tensors for the
+  `qwen35` / `qwen35moe` arch family ("MTP tensors are not used at
+  inference yet"), so the bundled GGUF and the unsloth GGUFs ship with
+  851 tensors and no MTP head. llama.cpp's MTP support (PR #22673,
+  merged 2026-05-16) currently covers other architectures only;
+  tracking that PR's follow-up work for when qwen35 / qwen35moe
+  consumer support lands. (Earlier README versions claimed MTP was
+  available without this caveat — confirmed empirically via
+  `gguf.GGUFReader` on both this bundle and `unsloth/Qwen3.6-27B-GGUF`,
+  2026-05-19.)
 ### Stamp choice