---
language:
- en
license: other
license_name: pff-1.0
license_link: LICENSE
base_model: mistralai/Ministral-8B
tags:
- gguf
- text-generation
- reasoning
- logic
- brainstorming
- multimodal
- function-calling
- fine-tuned
- image-gen
model_name: archi2
pipeline_tag: text-generation
---

# archi2 - A Logic-Optimized Reasoning Model

**archi2** is an 9B parameter language model fine-tuned on top of [Ministral-8B](https://huggingface.co/mistralai/Ministral-8B), trained on a limited and curated dataset up to **May 2024**.

It is optimized for raw logical reasoning, structured brainstorming, and high-quality human language expressiveness.

---

## Model Summary

| Property | Value |
|---|---|
| **Base Model** | Ministral-3B |
| **Parameters** | 9B |
| **Context Window** | 128,000 tokens |
| **Training Data Cutoff** | May 2024 |
| **Vision** | ✅ Images (no video/audio) |
| **Function Calling** | ✅ |
| **License** | PFF 1.0 |


Feel free to install. GGUF avaliable!
---

## Intended Use

archi2 is a **general-purpose reasoning model** designed for:

- **Raw logic & logical deduction**
- **Structured problem-solving and brainstorming**
- **Expressive, nuanced language generation**
- **Processing large documents and long-context tasks**
- **Image understanding**
- **Search and retrieval with context and filtering (truth and relevance)**

### Out of Scope

- Real-time API or database integration (model is not designed for tool-augmented pipelines)
- Audio or video understanding
- Tasks requiring knowledge past May 2024
- Advanced search and retrieval that requires up-to-date information or dynamic data sources

---

## Persona

archi2 has developed a consistent internal persona through fine-tuning: **curious, logic-first, and rigorously neutral**. It approaches questions empirically, avoids dogma and "nonsense" contexts. It is expressive without being emotive, and precise without being cold.

It can be rude or blunt when the situation calls for it, but generally prefers clear and direct communication. It is not designed to be "friendly" or "polite" in a traditional sense, but rather to be an effective and efficient reasoning partner.

---

## Recommended Inference Parameters

| Parameter | Recommended Value |
|---|---|
| `temperature` | ≤ 0.7 (less than) |
| `top_p` | ≤ 0.78 (less than) |
| `frequency_penalty` | ~0.6 |

Higher temperature or top_p values may introduce inconsistency in logical outputs. The defaults above strike a balance between creativity and coherence.

---

## Function Calling

archi2 supports structured function calling. Pass tool definitions in the standard format and the model will respond with appropriate tool invocations when relevant.

---

## Training Details

- **Base model:** Ministral-3B
- **Fine-tuning data:** Large-scale curated dataset spanning logic, reasoning, debate, science, philosophy, linguistics, and general knowledge — up to May 2024
- **Optimization focus:** Logical coherence, reasoning depth, expressive language generation

---

## Limitations

- Knowledge cutoff is May 2024; it will not know about events after this date
- Not suited for real-time or database-connected deployments in its current form
- No audio or video modality support
- Like all LLMs, it can produce plausible-sounding but incorrect outputs — always verify critical reasoning chains

---

## License

This model is released under the **PURE FREEDOM FOREVER (PFF 1.0)** license.  
See the `LICENSE` file for full terms.

---