🧬 Myth 4B

Name: Myth-4B
Availability: InStock

Myth 4B is a fusion of Qwen 3.5 4B and Qwen3.5-4B-Neo via SLERP (Spherical Linear Interpolation), built with a custom fusion engine that operates directly on safetensors — no mergekit required.

"Three forces, one entity — Myth"

Model Details

Architecture

Attribute	Value
Base Model	Qwen 3.5 4B
Parameters	4B
Context Length	32K
Architecture	`qwen3_5`
Format	FP16 (sharded, 8 files) + GGUF Q4_K_M
Fusion Method	SLERP (50% Qwen 3.5 + 50% Neo)

Lineage

Qwen 3.5 4B (multimodal base)
        |
        ├── SLERP ──→ 🧬 Myth 4B
        |
Qwen3.5-4B-Neo (reasoning expert)

Myth inherits:

Deep reasoning from Neo's reasoning-focused fine-tuning
Multimodal capabilities from Qwen 3.5's native architecture
Coding & math from Qwen 3.5's strong foundation
1M-token capable via Qwen 3.5's extended context architecture

Fusion Engine

Unlike most merges that use mergekit, Myth 4B was merged using a custom-built fusion engine:

myth_fusion.py — Custom SLERP Engine
  • Reads safetensors directly (no transformers dependency)
  • Processes one tensor at a time (low memory footprint)
  • Saves in shards (8 shards × ~100 tensors each)
  • Supports any model architecture
  • No GPU required

738 tensors were merged in 2 stages, with shard-based saving to prevent OOM on 8GB RAM hardware.

Files

File	Size	Description
`model-*.safetensors` (×8)	8.8 GB total	FP16 model weights (sharded)
`Myth-4B-Q4_K_M.gguf`	2.78 GB	Quantized GGUF (Q4_K_M)
`myth_system_prompt.md`	5 KB	Custom system prompt (Claude Fable 5-inspired)

Usage

Transformers (Python)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "dracko14/Myth-4B",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "dracko14/Myth-4B",
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": "You are Myth, a helpful AI assistant."},
    {"role": "user", "content": "Write a Python function for binary search."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

llama.cpp / Ollama

bash

# Download GGUF
wget https://huggingface.co/dracko14/Myth-4B/resolve/main/Myth-4B-Q4_K_M.gguf

# Run with llama.cpp
./llama-cli -m Myth-4B-Q4_K_M.gguf -p "Hello" -n 256

# Or with Ollama (create Modelfile first)
echo "FROM ./Myth-4B-Q4_K_M.gguf" > Modelfile
ollama create myth -f Modelfile
ollama run myth

System Prompt

Myth includes a custom system prompt inspired by Claude Fable 5: 📄 myth_system_prompt.md

Benchmarks

Benchmarks coming soon. Myth 4B inherits Qwen 3.5 4B's strong baseline with enhanced reasoning from Neo.

Training & Merge Details

Detail	Value
Compute	8GB RAM VPS (CPU only)
Merge Time	~15 minutes (738 tensors)
Quantization	llama.cpp (162 seconds)
Upload	~10 minutes
Total Cost	$0 (free VPS + free HuggingFace)

Limitations

Base model knowledge cutoff applies
Performance on CPU may be slow (4B params)
Not fine-tuned for specific downstream tasks
May exhibit base model biases

License

MIT — open for all use cases.

dracko14/
Myth-4B