MiniMythos-9B

MiniMythos-9B

Self-reliant coding & cybersecurity model with a fable-inspired system prompt. Qwen3.5 architecture, 1M context. Created by NilHRH.

Quick Start

GGUF (LM Studio / Ollama / llama.cpp)

Download the Q4_K_M GGUF from the repo releases and use it directly:

# llama.cpp example
./llama-cli -m MiniMythos-9B-Q4_K_M.gguf \
  --temp 0.6 --top-p 0.95 --top-k 20 \
  --prompt "<|im_start|>user\nWrite a Python one-liner palindrome checker.<|im_end|>\n<|im_start|>assistant\n<think>"

Transformers (requires base model weights)

from transformers import AutoModelForImageTextToText, AutoTokenizer

MODEL = "NilHRH/MiniMythos-9B"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    "NilHRH/MiniMythos-9B",
    config=MODEL,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Write a Python one-liner palindrome checker."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.6, top_p=0.95, top_k=20, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Benchmarks

vs Frontier Models

* MMLU with --limit 100 per subject (57 subjects). Full-eval numbers would be higher.

Local Inference (RTX 5060 Ti, 4-bit)

• Average speed: ~5 tok/s on 4-bit quantized Qwen3.5 architecture
• Covers code, math, reasoning, cybersecurity, and knowledge domains
• Full benchmark results in benchmark_results.json

System Prompt

MiniMythos uses a self-reliant fable-inspired system prompt baked into the chat template. Key traits:

• Self-reliance: Solves problems directly — no delegation to sub-agents or other models
• Lead with outcome: First sentence answers what happened or was found
• Progress verification: Audits claims against actual results before reporting
• Autonomy: Operates without real-time supervision; pauses only for destructive actions, scope changes, or blocked tasks
• Context awareness: Does not stop prematurely due to perceived context limits

Details

• Architecture: Qwen3.5-9B with 1M context (YaRN rope-scaled)
• Training: None — config-only modification (chat template + system prompt identity)
• Files: config.json, tokenizer.json, chat_template.jinja, MiniMythos-9B-Q4_K_M.gguf