Instructions to use GestaltLabs/Ornstein-Hermes-3.6-27b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use GestaltLabs/Ornstein-Hermes-3.6-27b with Transformers:

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

pipe = pipeline("image-text-to-text", model="GestaltLabs/Ornstein-Hermes-3.6-27b")
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 AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("GestaltLabs/Ornstein-Hermes-3.6-27b")
model = AutoModelForImageTextToText.from_pretrained("GestaltLabs/Ornstein-Hermes-3.6-27b")
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?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

HERMES

How to use GestaltLabs/Ornstein-Hermes-3.6-27b with HERMES:

# No code snippets available yet for this library.

# To use this model, check the repository files and the library's documentation.

# Want to help? PRs adding snippets are welcome at:
# https://github.com/huggingface/huggingface.js

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use GestaltLabs/Ornstein-Hermes-3.6-27b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "GestaltLabs/Ornstein-Hermes-3.6-27b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "GestaltLabs/Ornstein-Hermes-3.6-27b",
		"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/GestaltLabs/Ornstein-Hermes-3.6-27b

SGLang

How to use GestaltLabs/Ornstein-Hermes-3.6-27b 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 "GestaltLabs/Ornstein-Hermes-3.6-27b" \
    --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": "GestaltLabs/Ornstein-Hermes-3.6-27b",
		"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 "GestaltLabs/Ornstein-Hermes-3.6-27b" \
        --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": "GestaltLabs/Ornstein-Hermes-3.6-27b",
		"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"
						}
					}
				]
			}
		]
	}'

Unsloth Studio new

How to use GestaltLabs/Ornstein-Hermes-3.6-27b 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 GestaltLabs/Ornstein-Hermes-3.6-27b 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 GestaltLabs/Ornstein-Hermes-3.6-27b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for GestaltLabs/Ornstein-Hermes-3.6-27b to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="GestaltLabs/Ornstein-Hermes-3.6-27b",
    max_seq_length=2048,
)

Docker Model Runner
How to use GestaltLabs/Ornstein-Hermes-3.6-27b with Docker Model Runner:
```
docker model run hf.co/GestaltLabs/Ornstein-Hermes-3.6-27b
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Ornstein-hermes-3.6-27b

A Hermes-format function-calling fine-tune of Ornstein-3.6-27B, built on top of Qwen 3.6 27B's hybrid linear + full attention multimodal architecture. Trained on the top-quality Hermes-format slice of DJLougen/Acta, a curated 8-factor agentic tool-use dataset.

GGUF quantizations available at GestaltLabs/Ornstein-Hermes-3.6-27b-GGUF — Q8_0 down through aggressive 2-bit I-quants.

Support This Work

I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.

Support on Ko-fi

Intended Use

This model is specialized for Hermes-format function calling and agentic tool use:

<tools>[...]</tools> system-prompt tool registration
<think>...</think> reasoning blocks before tool selection
<tool_call>{...}</tool_call> structured function-call emission
<tool_response>...</tool_response> multi-turn tool-result handling

The base Ornstein-3.6-27B's vision encoder is preserved — image + text inputs still work — but training data was text-only, so vision-language behavior is inherited from the base, not tuned.

Training Details


Base model	GestaltLabs/Ornstein-3.6-27B
Dataset	DJLougen/Acta — `hermes_reasoning` ∪ `hermes_multiturn`, filtered to `quality_score >= 0.85`
Training rows	1,559 train / 153 validation
Method	LoRA (r=32, α=32, dropout=0.0) on language layers only — attention + MLP linears
Trainable params	233.5M (0.85% of 27.6B)
Precision	bf16 (no quantization during training)
Sequence length	4,096 tokens (dataset p99 ≈ 2,400)
Optimizer	AdamW, lr 1e-4, cosine schedule, 3% warmup
Effective batch size	16 (per-device 2 × grad-accum 8)
Epochs	2 (196 optimizer steps)
Hardware	1× NVIDIA RTX PRO 6000 Blackwell (97 GB)
Wall time	≈ 2h 03m
Framework	Unsloth 2026.4.8 + TRL 0.24 + PEFT 0.19

Eval Loss Trajectory

Eval every 25 steps on a held-out 153-sample Hermes validation set:

step	eval_loss
25	0.4809
50	0.4433
75	0.4269
100	0.4137
125	0.4043
150	0.3991
175	0.3957
200	0.3953 ← best

Monotonic descent across all 8 evaluation points, no overfitting signal. load_best_model_at_end=True, so the saved adapter (and this merge) is the best-eval checkpoint.

Why this filter?

The Acta dataset mixes 5 source corpora at varying quality. The Hermes subsets (hermes_reasoning, hermes_multiturn) ship with the canonical role layout (system / user / assistant / tool) that Qwen's chat template handles natively, and have the highest mean quality scores in the dataset (0.902 / 0.833 respectively). The other sources (flamefox_agentic, smolagents_code, swe_agent_glm) either use atypical role assignment or score below 0.85 and were excluded to keep the training distribution clean.

Usage

Transformers

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "GestaltLabs/Ornstein-hermes-3.6-27b"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True
)

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a city.",
            "parameters": {
                "type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"],
            },
        },
    }
]

messages = [{"role": "user", "content": "What's the weather in Tokyo?"}]
inputs = processor.apply_chat_template(
    messages, tools=tools, add_generation_prompt=True,
    tokenize=True, return_tensors="pt"
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[-1]:], skip_special_tokens=True)[0])

llama.cpp / Ollama / LM Studio (via GGUF)

See the GGUF repo — pick a quant that fits your memory:

Q4_K_M — strong default for 24 GB cards
Q5_K_M / Q6_K — higher fidelity if you have 32–48 GB
Q8_0 — near-lossless, ~28 GB
IQ4_NL / IQ4_XS — imatrix-calibrated 4-bit, smaller than Q4_K_M
IQ3_M / IQ2_M — aggressive but usable for tight VRAM budgets

Architecture

Inherited unchanged from Ornstein-3.6-27B:

Qwen3_5ForConditionalGeneration — Qwen 3.6 dense with linear + full attention interleaved (Gated Delta Net) + vision encoder
~27B parameters total (dense, multimodal)
Hidden size 5,120 / 64 layers
Attention: 24 heads, 4 KV heads, head_dim 256, full-attention every 4 layers (linear otherwise)
Context length: 262,144 tokens

License

Apache 2.0 — inherited from the Qwen 3.6 base release. Acta is also Apache 2.0.

Citation

If you use this model, please consider citing the dataset:

@dataset{lougen_acta_2026,
  author = {DJLougen},
  title = {Acta: A Premium Curated Sample of High-Quality Agentic Tool-Use Conversations},
  year = {2026},
  url = {https://huggingface.co/datasets/DJLougen/Acta}
}