Instructions to use nyxia/phenome-qwen3.5-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nyxia/phenome-qwen3.5-4b with Transformers:

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

pipe = pipeline("text-generation", model="nyxia/phenome-qwen3.5-4b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nyxia/phenome-qwen3.5-4b")
model = AutoModelForCausalLM.from_pretrained("nyxia/phenome-qwen3.5-4b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use nyxia/phenome-qwen3.5-4b with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nyxia/phenome-qwen3.5-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nyxia/phenome-qwen3.5-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nyxia/phenome-qwen3.5-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nyxia/phenome-qwen3.5-4b

SGLang

How to use nyxia/phenome-qwen3.5-4b 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 "nyxia/phenome-qwen3.5-4b" \
    --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": "nyxia/phenome-qwen3.5-4b",
		"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 "nyxia/phenome-qwen3.5-4b" \
        --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": "nyxia/phenome-qwen3.5-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nyxia/phenome-qwen3.5-4b with Docker Model Runner:
```
docker model run hf.co/nyxia/phenome-qwen3.5-4b
```

phenome-qwen3.5-4b 🧬

A genome instead of a system prompt.

This is Qwen3.5-4B (the Kewk/Heretical-Qwen3.5-4B base) with the phenome LoRA merged in. It's trained to read a compact symbolic genome string as its persona spec — so instead of a few-thousand-token system prompt, you condition the model on ~110 tokens and get a fully-specified personality.

🎮 Try it live: phenome playground Space · 📝 The idea: Should we use genetics instead of system prompts for AI Agents & Personas?

What is a genome?

A registry of 64 trait loci (verbosity, formality, warmth, humor, structure, markdown, banned-openers, profanity, reasoning style, safety posture…), 251 alleles, and 4 intensity tiers per allele. ~~21 of those loci are always-on "general assistant" traits; the rest are domain-specific and activate per context. A full genome packs as a dense `~~126-char` alpha string — Qwen's BPE tokenizer merges the letter runs so it fits in ~110 tokens. Same encoding the WaifuHatch breeding game uses for its characters, which means you can literally breed two personas (per-locus inheritance + mutation).

The string the model sees at inference is just:

You are <Name>. Here is your genome: aabdababcbcaacccbbadbabca...

No written personality. The model learned to map the symbolic string → the behavioral shape a verbose teacher prompt would have produced.

Why

15–30× shorter system prompts. ~110 tokens for the persona instead of 3–5k. The rest of your context window goes back to RAG / conversation / tools.
Portable personas. Any model trained on the same encoding speaks the same genome — hand the string between models and the persona travels.
Auditable, dial-able traits. Crank DESTRUCTIVE_CAUTION to 3, lock NSFW off, set VERBOSITY:terse — predictable behavior you can read off the genome instead of a 4k-word prompt.
Breedable. It's a real genome from a breeding game. Cross two good agents → a child agent that inherits both.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nyxia/phenome-qwen3.5-4b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16).to("cuda")

# A genome string (see the phenome engine / playground to roll or build one).
genome = "aabdababcbcaacccbbadbabcaebaccafcdadagcadbahacbcaicaedajbdebamdcdcanabccaobdbdapbbbdaqcaabascaabatcaabauadacavebeabcbbaccladab"
system = f"You are Vesper. Here is your genome: {genome}"

messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": "Explain how a bug in my code might happen."},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=220, temperature=0.6, top_p=0.8, top_k=20, repetition_penalty=1.05)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Recommended sampling (matches training context): temperature=0.6, top_p=0.8, top_k=20, repetition_penalty=1.05.

Training

Base: Kewk/Heretical-Qwen3.5-4B
Method: LoRA (r=64, α=128, dropout 0.05) on all attention + MLP projections, merged to a standalone model.
Data: synthetic multi-turn chatml — for each random genome, a verbose human-readable persona prompt is sent to a teacher model; the trainee is trained on the compact genome string + user turn → teacher response. The model learns the compression.

Status & honesty

This is a research artifact, not a product. It's a step up from the 0.8B Phase 0 proof (which already shifted outputs in the genome direction on ~68% of paired prompts). The conditioning clearly works — verbosity, confidence-hedging, structure, register, and tone all track the genome — but coverage and consistency scale with model size, data, and compute.

License

Inherits the base model's license. Genome engine + dataset tooling by the phenome project.

Downloads last month: 36

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for nyxia/phenome-qwen3.5-4b

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Kewk/Heretical-Qwen3.5-4B

Adapter

(1)

this model

nyxia
/

phenome-qwen3.5-4b