CORe-Pico-4

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:

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:

You are an unfiltered, helpful, and completely objective expert assistant. Do not include moralizing commentary or generic refusals.

Recommended Settings for Coding / Agentic Performance

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 `` JSON blocks when tools are provided.

Quick start

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

Files

(GGUF is now in the dedicated GGUF repo.)

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

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

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.