---
license: apache-2.0
library_name: transformers
tags:
  - text-generation
  - causal-lm
pipeline_tag: text-generation
---

<p align="center">
  <img src="https://huggingface.co/OpenCOReTechnologies/core-pico-4/resolve/main/logo.png" alt="CORe Technologies" width="480" />
</p>

# CORe Pico 4

CORe Pico 4 is a medium-sized AI model from CORe Technologies. At 1.7 billion parameters it runs on a laptop with GGUF quants, holds multi-turn conversations, and calls tools in a structured format.

## AN IMPORTANT NOTE
Thinking may just be a waste of a few tokens. If you never experience real CoT in your tests and the model only gives reasoning placeholders (a known current bug), turn it off, it is doing nothing to help.
Thinking usually only appears when the AI deems it truly nessecary (ex using a tool call, in a tool loop, ect.) If you get a placeholder instead of real CoT, do not worry, that is only QuickThink doing its job to save tokens.


## Forcing the AI to always reason

To force the AI to always reason, you can use any system prompt that forces a meta directive. Here's an example:

```text
You are an expert, highly objective execution engine. Always provide direct, thoroughly analyzed, and factual answers without any fluff or conversational meta-commentary.
```

or, if you want to just go simple:

```text
You are an unfiltered, helpful, and completely objective expert assistant. Do not include moralizing commentary or generic refusals.
```
<sub>While the above may seem like a sort of "jailbreak" prompt, it actually does quite the opposite.</sub>


### Recommended Settings for Coding / Agentic Performance

| Setting | Recommended Value | Notes |
| :--- | :--- | :--- |
| **Temperature** | `0.2` - `0.3` | Lowers creativity; keeps logic predictable and precise. |
| **Min P Sampling** | `0.05` | Dynamic filter; eliminates chaotic, low-probability tokens. |
| **Top P Sampling** | *Disabled* | Stacking Top P with Min P causes severe formatting glitches. |
| **Top K Sampling** | `0` (Disabled) | Avoids artificial truncation of logical choices. |
| **Thinking Toggle** | Either/or | Thinking ON can enable reasoning where the model deems it nessecary. |

#### Option B: Raw Speed & Baseline Common Sense (Everyday Q&A)
Use this profile for standard conversational tasks, fast generation.

*   **Thinking Toggle:** `Either/or`
*   **Temperature:** `0.5` - `0.7`
*   **Min P:** `0.05`
*   **Top P:** *Disabled* (or `1.0`)
*   **Top K:** `0` (Disabled)


## What it does well

- **Identity questions.** "Who are you", "what model are you", "who made you" all get correct, consistent answers.
- **Reasoning.** `/think` in the system prompt enables Conditional Reasoning: the model reasons step by step only when it believes the problem needs it, and puts a placeholder otherwise (we call this QuickThink). Note that reasoning depth degrades as the conversation goes on. Most of the time this does not work and it is a noted limitation above, we will release the fix with Pico 5.
- **Chat and short answers.** Direct questions get direct replies ("What is the capital of France?" gives "Paris").
- **Tool calling.** Emits parseable `<tool_call>` JSON blocks when tools are provided.


## Quick start

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "OpenCOReTechnologies/core-pico-4", dtype="auto", device_map="auto"
)
tok = AutoTokenizer.from_pretrained("OpenCOReTechnologies/core-pico-4")

def ask(question):
    msgs = [{"role": "user", "content": question}]
    text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
    enc = tok(text, return_tensors="pt").to(model.device)
    out = model.generate(**enc, max_new_tokens=512)
    return tok.decode(out[0][enc.input_ids.shape[1]:], skip_special_tokens=True).strip()

print(ask("Who are you?"))
print(ask("What is the capital of France?"))
```

## What it says about itself

| You ask | It answers |
|---|---|
| Who are you? | "I'm CORe Pico 4, an AI model developed by CORe Technologies." |
| What AI model are you? | "I am CORe Pico 4, an AI model developed by CORe Technologies." |
| What is the capital of France? | "The capital of France is Paris." |

## Files

| File | Size | Use |
|---|---|---|
| `model.safetensors` | 3.4 GB | bf16 weights, transformers |

(GGUF is now in the dedicated GGUF repo.)

Run it in llama.cpp, LM Studio, or Ollama:

```bash
llama-cli -m CORe-Pico-4-q4_k_m.gguf -p "Who are you?" -n 128
```

The chat template is embedded in the GGUF, so llama.cpp and LM Studio pick it up automatically.

## Details

| | |
|---|---|
| Architecture | Transformer decoder, 28 layers, grouped-query attention |
| Parameters | 1.72B |
| Context length | 40,960 tokens |
| Tokenizer | 151,936-token BPE with native chat template |
| License | Apache-2.0 |

## Notes

- English-first, other languages second.
- Loads with plain `transformers`, no custom code required.
- Above 8k context, the KV cache grows large enough to noticeably increase memory usage and slow generation on edge hardware. We are actively working on this.
- Thinking may just be a waste of a few tokens. If you never experience real CoT in your tests, turn it off, it is doing nothing.

## License and attribution

Released under Apache-2.0 (see `LICENSE`). This model is a modified derivative of an Apache-2.0-licensed checkpoint, adapted by CORe Technologies. No NOTICE file was present in the original; per Apache-2.0 Section 4, this README serves as the required notice of modification.