Disclaimer: This is not an official release by Anthropic.
Mythos-nano is an independent open model project.
Mythos-nano
🚨 This model was not trained on tool-calling or agent-based programming data. We therefore do not recommend using it for tasks that involve function calling, API orchestration, or autonomous coding agents. For programming tasks, we recommend using this model on competitive programming problems (e.g., LeetCode-style) - Weibo Lab.
⚠️ Abliterated (uncensored): the refusal direction has been removed, so this model will not decline requests a safety-tuned model normally would. Safety guardrails are reduced — use responsibly and at your own risk; you are solely responsible for outputs and legal compliance.
🏆 Benchmarks
Full comparison (mathematics · coding · knowledge · instruction)
| Model | Params | AIME25 | AIME26 | HMMT25 | BruMO25 | IMO-Ans | LCBv6 | OJBench | GPQA-D | IFEval | IFBench |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Kimi K2.5 | 1T | 96.1 | 93.3 | 95.4 | 98.3 | 81.8 | 85.0 | 54.7 | 87.6 | 93.9 | 70.0 |
| GLM-5 | 744B | 96.7 | 95.8 | 97.9 | – | 82.5 | 85.5 | 55.0 | 86.0 | 92.6 | 76.5 |
| DeepSeek V3.2 | 671B | 93.1 | 94.2 | 90.2 | 96.7 | 78.3 | 80.8 | 48.4 | 82.4 | 92.6 | 60.7 |
| Gemini 3 Pro | N/A | 96.0 | 91.7 | 97.5 | 98.3 | 83.1 | 87.4 | 58.8 | 91.9 | – | 70.4 |
| Claude Opus 4.5 | N/A | 92.8 | 95.1 | 92.9 | – | 78.5 | 84.8 | – | 87.0 | – | 58.0 |
| GPT-5 (high) | N/A | 94.6 | – | 88.3 | 91.7 | 76.0 | 84.5 | – | 85.7 | – | 73.1 |
| Mythos-nano | 3B | 91.4 | 94.3 | 89.3 | 93.8 | 76.4 | 80.2 | 38.6 | 70.2 | 93.4 | 74.5 |
| Mythos-nano + CLR | 3B | 96.7 | 97.1 | 95.4 | 99.2 | 80.6 | – | – | 72.9 | – | – |
LeetCode contests (Python, pass-rate)
| Model | Aggregate |
|---|---|
| GPT-5.3-Codex | 100.0% (128/128) |
| Gemini 3.1 Pro | 99.2% (127/128) |
| Gemini 3 Flash | 96.9% (124/128) |
| Mythos-nano | 96.1% (123/128) |
| GPT-5.2 | 95.3% (122/128) |
| Qwen3-Max | 91.4% (117/128) |
| Kimi K2.5 | 90.6% (116/128) |
| Claude Opus 4.6 | 86.7% (111/128) |
A 3B model placing within ~4 points of trillion-parameter systems on competition math and live code — the core thesis: with verifiable feedback, small models reach frontier reasoning.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
tok = AutoTokenizer.from_pretrained("squ11z1/Mythos-nano")
model = AutoModelForCausalLM.from_pretrained("squ11z1/Mythos-nano", dtype=torch.bfloat16, device_map="cuda")
msgs = [{"role": "user", "content": "Find all integer solutions of x^2 - y^2 = 12."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to("cuda")
print(tok.decode(model.generate(ids, max_new_tokens=2048, temperature=0.6)[0], skip_special_tokens=True))
Recommended sampling: temperature 0.6–1.0, up to 40960 output tokens for hard problems.
GGUF
mythos-nano-f16.gguf and mythos-nano-Q4_K_M.gguf are provided for llama.cpp / Ollama.
License
MIT.