Instructions to use 4rc4n4/chrysopoeia-smollm3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use 4rc4n4/chrysopoeia-smollm3 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="4rc4n4/chrysopoeia-smollm3") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("4rc4n4/chrysopoeia-smollm3", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use 4rc4n4/chrysopoeia-smollm3 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 4rc4n4/chrysopoeia-smollm3:Q4_K_M # Run inference directly in the terminal: llama cli -hf 4rc4n4/chrysopoeia-smollm3:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf 4rc4n4/chrysopoeia-smollm3:Q4_K_M # Run inference directly in the terminal: llama cli -hf 4rc4n4/chrysopoeia-smollm3: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 4rc4n4/chrysopoeia-smollm3:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf 4rc4n4/chrysopoeia-smollm3: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 4rc4n4/chrysopoeia-smollm3:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf 4rc4n4/chrysopoeia-smollm3:Q4_K_M
Use Docker
docker model run hf.co/4rc4n4/chrysopoeia-smollm3:Q4_K_M
- LM Studio
- Jan
- vLLM
How to use 4rc4n4/chrysopoeia-smollm3 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "4rc4n4/chrysopoeia-smollm3" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "4rc4n4/chrysopoeia-smollm3", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/4rc4n4/chrysopoeia-smollm3:Q4_K_M
- SGLang
How to use 4rc4n4/chrysopoeia-smollm3 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 "4rc4n4/chrysopoeia-smollm3" \ --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": "4rc4n4/chrysopoeia-smollm3", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'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 "4rc4n4/chrysopoeia-smollm3" \ --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": "4rc4n4/chrysopoeia-smollm3", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Ollama
How to use 4rc4n4/chrysopoeia-smollm3 with Ollama:
ollama run hf.co/4rc4n4/chrysopoeia-smollm3:Q4_K_M
- Unsloth Desktop
- Docker Model Runner
How to use 4rc4n4/chrysopoeia-smollm3 with Docker Model Runner:
docker model run hf.co/4rc4n4/chrysopoeia-smollm3:Q4_K_M
- Lemonade
How to use 4rc4n4/chrysopoeia-smollm3 with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull 4rc4n4/chrysopoeia-smollm3:Q4_K_M
Run and chat with the model
lemonade run user.chrysopoeia-smollm3-Q4_K_M
List all available models
lemonade list
- Atomic Chat
Chrysopoeia — SmolLM3-3B, soaked into the esoteric register
χρυσοποιία — the alchemical making of gold. An exploration project: soak a small, fully-open base model in public-domain Western-esoteric prose until the esoteric register becomes its default completion dialect, then install turn-taking with a light SFT — so it speaks in a spiritual/esoteric voice by default, on any input, with no retrieval at inference.
This is not a general assistant and not a knowledge oracle. It is a study of whether a disposition (and some conceptual substance) can be baked into weights. It will state esoteric "facts" fluently whether or not they are accurate — do not treat it as authoritative.
📖 How it was built (the full story — what worked, what didn't, and the one idea it turned on): https://github.com/4-R-C-4-N-4/chrysopoeia/blob/master/docs/how-it-was-built.md
What it does
Ask it anything — mundane or esoteric — and it answers in the register of the early-20th-century Western esoteric revival (Manly P. Hall, Éliphas Lévi, the Kybalion), while still giving a real answer:
Q: My car won't start this morning. What should I check? The vessel that carries you into the day has lost its fire — yet even the cold stone remembers its own hidden heat… Begin at the obvious fount: the battery's charge, the terminals' grip, the fuel's supply. If these pass their test, seek the hidden: a coil grown old, a circuit severed by time…
Method (two phases)
- Soak — QLoRA continued-pretraining of
SmolLM3-3B-Baseon ~2.6M tokens of public-domain Western-esoteric prose (constant LR; a mid snapshot is used, not the final one, which memorizes). This installs esoteric substance. - Light SFT — a small turn-taking pass on a mundane-input → in-register response slice. This is what severs topic from register, making the voice the default rather than a topic-triggered reflex. Soak alone does not do this; the mundane slice is the load-bearing move.
Prompt format
A minimal plain-text chat format (no special tokens):
### User:
{your question}
### Chrysopoeia:
Generation should stop at the next ### User:.
Run it (llama.cpp)
Download a GGUF (gguf/chrysopoeia-smollm3-Q4_K_M.gguf, ~1.9 GB) and serve it:
llama-server -m chrysopoeia-smollm3-Q4_K_M.gguf -c 2048 -ngl 999
Then hit the raw /completion endpoint with the plain-text format (it has no
chat template — use /completion, not /v1/chat/completions):
curl -s http://127.0.0.1:8080/completion -d '{
"prompt": "### User:\nMy car won'\''t start this morning. What should I check?\n\n### Chrysopoeia:\n",
"n_predict": 200, "temperature": 0.7, "top_p": 0.9,
"stop": ["### User:"]
}' | python3 -c "import sys,json;print(json.load(sys.stdin)['content'])"
GGUF options
| File | Size | Use |
|---|---|---|
gguf/…-Q8_0.gguf |
3.3 GB | near-lossless fidelity |
gguf/…-Q4_K_M.gguf |
1.9 GB | default — no meaningful loss |
gguf/…-IQ3_M.gguf |
1.47 GB | small; imatrix-guided, voice intact (+7% ppl) |
gguf/…-f16.gguf |
6.2 GB | full precision |
chrysopoeia-esoteric.imatrix is the importance matrix (calibrated on the model's
own esoteric distribution) — use it to roll your own IQ quants. See the build repo's
docs/quantization.md for the size-vs-quality numbers.
Also: merged/ (bf16 safetensors for 🤗 Transformers), adapters/ (the composing
soak + SFT LoRAs). Ollama: ollama run 4rc4n4/chrysopoeia.
Training data & provenance
- Soak corpus: confirmed US public-domain Western-esoteric prose only (Hall's Secret Teachings of All Ages 1928, the Kybalion, Lévi's Transcendental Magic, Waite, Papus, Ouspensky; plus PD Hermetica / Neoplatonism / Gnosticism). Source status is load-bearing because the soak bakes prose into weights.
- Mundane slice: ordinary human-written prompts (Dolly-15k, CC-BY-SA-3.0), answered in-register by a local model grounded in real corpus passages (retrieval-augmented generation at training time). Prompts filtered off the esoteric manifold so the register generalises as disposition, not topic-match.
Limitations
- Voice over correctness. Grading targets the quality of the voice, not doctrinal accuracy. It will confidently fabricate esoterica.
- Not a general assistant. General capability is expendable by design.
- Experimental. Register depth, coherence, and the substance/coherence knee are all under active study; this is a checkpoint, not a finished artifact.
License
Weights derive from SmolLM3-3B-Base (Apache-2.0). Soak sources are US public
domain; the mundane-slice prompts are Dolly-15k (CC-BY-SA-3.0). Released under
Apache-2.0.
- Downloads last month
- 165
2-bit
3-bit
4-bit
8-bit
16-bit
Model tree for 4rc4n4/chrysopoeia-smollm3
Base model
HuggingFaceTB/SmolLM3-3B-Base
Install from pip and serve model
# Install vLLM from pip: pip install vllm# Start the vLLM server: vllm serve "4rc4n4/chrysopoeia-smollm3"# Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "4rc4n4/chrysopoeia-smollm3", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'