File size: 4,167 Bytes
6222150 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | ---
license: mit
language:
- en
pipeline_tag: text-generation
tags:
- reasoning
- math
- code
- qwen2
- mythos-nano
base_model:
- WeiboAI/VibeThinker-3B
base_model_relation: finetune
---
</a>
> **Disclaimer:** This is **not** an official release by Anthropic.
> Mythos-nano is an independent open model project.
# Mythos-nano
<blockquote style="border-left: 4px solid #ff6b6b; background-color: #fff5f5; padding: 10px 15px; margin: 10px 0; color: #cc3333;">
<span style="font-weight: bold;">🚨 </span> 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.
</blockquote>
<blockquote style="border-left: 4px solid #ff6b6b; background-color: #fff5f5; padding: 10px 15px; margin: 10px 0; color: #cc3333;">
<span style="font-weight: bold;">⚠️ </span> 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.
</blockquote>
## 🏆 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
```python
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. |