89 lines
4.2 KiB
Markdown
89 lines
4.2 KiB
Markdown
|
---
|
|||
|
|
license: mit
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- reasoning
|
|||
|
|
- math
|
|||
|
|
- code
|
|||
|
|
- qwen2
|
|||
|
|
- mythos-nano
|
|||
|
|
base_model:
|
|||
|
|
- WeiboAI/VibeThinker-3B
|
|||
|
|
base_model_relation: finetune
|
|||
|
|
---
|
|||
|
|
> Mythos-nano tool-calling is coming, but check out Merlin-Agent!
|
|||
|
|
https://huggingface.co/Merlin-Research/Merlin-Agent
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
</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.
|