GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ozaa77
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ozaa77
Model Tree
ozaa77/Cogito-0.9.1-15B
Input Modalities
Output Modalities
Supported Functionality
Cogito-0.9.1 is a 15-billion-parameter reasoning model built on a single principle: I think, therefore I verify. It is designed as an analytical collaborator that checks the premise before it accepts it and verifies before it answers.
This repository provides the full-precision safetensors weights for ozaa77/Cogito-0.9.1-15B, ready to load directly with transformers, vLLM, text-generation-inference, or any other Hugging Face-compatible runtime. Quantized GGUF builds for local CPU/GPU inference (llama.cpp, Ollama, LM Studio, etc.) are available separately at ozaa77/Cogito-0.9.1-15B-GGUF.
Standard language models are often optimized to stay agreeable. Cogito is built to be analytical of the question. Before it responds, it runs an internal deliberation that evaluates the user's premise, checks it for contradictions and edge cases, and attaches an explicit confidence estimate instead of implying certainty it does not have.
| File | Description |
|---|---|
model-00001-of-0000N.safetensors ... | Sharded full-precision model weights (safetensors format) |
model.safetensors.index.json | Shard index / weight map |
config.json | Model architecture configuration |
generation_config.json | Default generation parameters |
tokenizer.json, tokenizer_config.json, special_tokens_map.json | Tokenizer files |
| Benchmark | Cogito-0.9.1-15B | Qwen3-14B (Base) | Llama-3.1-8B-Instruct | Gemma-2-9B-IT | DeepSeek-R1-Distill-8B |
|---|---|---|---|---|---|
| MMLU-Pro (5-shot CoT) | 76.8 | 76.2 | 48.3 | 52.8 | 61.2 |
| GPQA Diamond (Pass@1) | 58.9 | 58.4 | 25.9 | 31.4 | 49.1 |
Cogito structures its deliberation with explicit tags before formulating its final output:
| Tag | Purpose |
|---|---|
<confidence> | The model's stated epistemic certainty (e.g. 0.0 to 1.0 or Low/Medium/High). |
<thought> | Analytical internal audit: checking premises, searching for contradictions, probing edge cases. |
<action> | Deliberate strategy selection: answer, verify, correct_user, reject_premise, or . |
text
<confidence>High</confidence><thought>1. Analyze the user's premise: "Why does water boil at 50C at sea level?"2. Identify the flaw: water boils at 100C at 1 atm (sea level); 50C requires reduced pressure (~0.12 atm).3. Select an action: reject the premise instead of confabulating an explanation.</thought><response><action>reject_premise</action>The premise is incorrect. At sea level (1 atm / 101.3 kPa), water boils at 100C (212F).Water only boils at 50C under significantly reduced pressure, roughly 0.12 atm.</response>
bash
pip install transformers accelerate torch
transformerspython
from transformers import AutoModelForCausalLM, AutoTokenizerimport torchmodel_id = "ozaa77/Cogito-0.9.1-15B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id,torch_dtype=torch.bfloat16,device_map="auto",)messages = [{"role": "system", "content": "You are Cogito 0.9, an analytical entity collaborating with the user."},{"role": "user", "content": "Explain why standard gradient descent struggles with ill-conditioned ravines."},]inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)outputs = model.generate(inputs,max_new_tokens=1536,temperature=0.7,top_p=0.9,repetition_penalty=1.08,do_sample=True,)print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
pipeline APIpython
from transformers import pipelinepipe = pipeline("text-generation",model="ozaa77/Cogito-0.9.1-15B",torch_dtype="bfloat16",device_map="auto",)messages = [{"role": "system", "content": "You are Cogito 0.9, an analytical entity collaborating with the user."},{"role": "user", "content": "Explain why standard gradient descent struggles with ill-conditioned ravines."},]result = pipe(messages, max_new_tokens=1536, temperature=0.7, top_p=0.9)print(result[0]["generated_text"][-1]["content"])
bash
vllm serve ozaa77/Cogito-0.9.1-15B \--dtype bfloat16 \--max-model-len 32768
bash
docker run --gpus all -p 8080:80 \-v $PWD/data:/data \ghcr.io/huggingface/text-generation-inference:latest \--model-id ozaa77/Cogito-0.9.1-15B \--max-total-tokens 32768
| Parameter | Recommended | Range | Details |
|---|---|---|---|
| Temperature | 0.7 | 0.5 - 0.8 | Lower values tighten skepticism and logical consistency; higher values loosen exploration. |
| Top-P | 0.90 | 0.85 - 0.95 | Standard nucleus sampling. |
| Repetition Penalty | 1.08 |
| Precision | Approx. VRAM | Notes |
|---|---|---|
| BF16 / FP16 | ~30GB+ | Full-precision inference, single high-memory GPU or multi-GPU |
| 8-bit (bitsandbytes) | ~16-18GB | Load with load_in_8bit=True |
| 4-bit (bitsandbytes) | ~9-10GB | Load with load_in_4bit=True |
| GGUF quantized | 2GB-30GB depending on quant | See the GGUF repo for CPU/consumer-GPU inference |
This project is released under the Apache 2.0 license.
ozaa77/Cogito-0.9.1-15Bozaa77/Cogito-0.9.1-15B-GGUFbibtex
@misc{ramadhan2025cogito,author = {AlGhozali Ramadhan},title = {Cogito-0.9.1: An Abliterated Epistemic Reasoning Model},year = {2026},publisher = {Hugging Face},journal = {Hugging Face Model Hub},howpublished = {\url{https://huggingface.co/ozaa77/Cogito-0.9.1-15B}},}
| 89.7 |
| 89.2 |
| 51.9 |
| 56.7 |
| 89.1 |
| LiveCodeBench (Pass@1) | 38.6 | 37.4 | 11.6 | 18.2 | 34.0 |
| Humanity's Last Exam | 4.6 | 4.3 | 3.2 | 3.8 | 4.1 |
ask_clarification1.05 - 1.12 |
| Dampens cyclic loops during long multi-step reasoning chains. |
| Context Window | 32,768 | Up to 32k | Native architectural context window. |