---
license: llama3.1
language:
- en
pipeline_tag: text-generation
library_name: transformers
inference: true
base_model:
- NousResearch/Hermes-3-Llama-3.1-8B
- NousResearch/DeepHermes-3-Llama-3-8B-Preview
- nvidia/Llama-3.1-Nemotron-Nano-8B-v1
- deepseek-ai/DeepSeek-R1-Distill-Llama-8B
tags:
- merge
- dare_ties
- llama
- llama-3.1
- mist
---

# MIST-1-8B

MIST-1-8B (formerly MIST-Mini) is the smallest and fastest model in the **MIST model family** by [olaverse](https://huggingface.co/olaverse). Built by blending 4 specialized Llama 3.1 8B models using DARE+TIES — delivering strong performance at maximum speed. 
fast, thorough, great for everyday use

## MIST Model Family

| Model | Params | Speed | Status |
|---|---|---|---|
| **MIST-1-8B** | 8B | ~63 tok/s | ✅ Available |
| [MIST-1-70B](https://huggingface.co/olaverse/MIST-1-70B) | 70B | ~23 tok/s | ✅ Available |
| [MIST-1-140B](https://huggingface.co/olaverse/MIST-1-140B) | 140B | ~8 tok/s | ✅ Available |

---

## Key Strengths

- ⚡ **Fastest** — 63 tok/s on H200, great for real-time applications
- 🧠 **Strong Reasoning** — DeepSeek R1 distillation
- 💻 **Clean Code** — production-ready with comments
- 📐 **Math** — accurate step-by-step solving
- 🤝 **Helpful** — low refusal rate
- 📦 **Lightweight** — 15GB, runs on consumer GPUs

---

## Benchmark Results

| Task | Speed | Quality |
|---|---|---|
| Reasoning | 4.5s | ✅ Correct |
| Coding | 4.0s | ✅ Clean code |
| Math | 4.0s | ✅ Step-by-step |
| General | 4.0s | ✅ Accurate |
| Instruction | 4.0s | ✅ Precise |

**Average: 63 tok/s**

---

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "olaverse/MIST-Mini-8B",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("olaverse/MIST-Mini-8B")

messages = [{"role": "user", "content": "Your question here"}]
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=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Hardware Requirements

| Precision | VRAM Required |
|---|---|
| bfloat16 | 16GB (RTX 3090/4090) |
| 4-bit | 6GB (RTX 3060+) |

---

## Recommended Generation Settings

These settings were verified through testing. Without `repetition_penalty` 
and `min_p` the model will ramble and not stop cleanly.

```python
outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    min_p=0.05,
    repetition_penalty=1.5,
    eos_token_id=[128040, 128009, 128001],
    pad_token_id=128001,
)
```

### Stop Tokens
This model's ChatML parents (`<|im_end|>`) survived the DARE+TIES merge 
alongside Llama 3.1 native tokens. Use all three:

| Token | ID | Source |
|---|---|---|
| `<\|im_end\|>` | 128040 | Hermes/Nemotron parents |
| `<\|eot_id\|>` | 128009 | Llama 3.1 native |
| `<\|end_of_text\|>` | 128001 | Llama 3.1 native |

## License

[Llama 3.1 Community License](https://llama.meta.com/llama3/license/)