Sky v2.0-Lite is the smallest and fastest member of the Sky v2.0 family. It combines K=2 step reduction (keeping only CREST Steps 1 and 2) with the option for 4-bit quantization at load time, making it possible to run a CREST adaptive-depth model on hardware as modest as an 8GB laptop GPU or even a Google Colab free tier T4.

Despite being the lightest variant, it still has adaptive depth — the model still decides per-token how much computation to use. It is not a standard transformer; it is a real CREST model.

Who is This For?

🎓 Students exploring adaptive-depth LLMs on a budget
💻 Developers building AI-powered applications on consumer hardware
📱 Edge/on-device deployment where every GB matters
🧪 Researchers who want to experiment with CREST without cloud GPUs

Model Details

Property	Value
Architecture	CREST (0labs proprietary)
Total Parameters	6.47B
CREST Steps (K)	2
Hidden Dimension	2,560
Layers	32
Attention Heads	32
Context Window	32,768 tokens
Disk Size	~13GB (BFloat16)
VRAM (BF16)	~13GB
VRAM (INT4)	~4GB
Vocabulary Size	248,320
License	Apache 2.0

Quick Start

Recommended: 4-Bit Loading (~4GB VRAM) ⭐

This is the recommended way to run Sky v2.0-Lite on consumer hardware:

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# 4-bit quantization config — uses only ~4GB VRAM
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
)

model = AutoModelForCausalLM.from_pretrained(
    "0labs-in/Sky-v2.0-Lite",
    trust_remote_code=True,       # Required — CREST is a custom architecture
    quantization_config=quantization_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("0labs-in/Sky-v2.0-Lite")

# Chat
prompt = "Explain quantum computing in simple terms."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Full Precision Loading (needs ~13GB VRAM)

python

model = AutoModelForCausalLM.from_pretrained(
    "0labs-in/Sky-v2.0-Lite",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Note: trust_remote_code=True is required — CREST uses custom architecture code included in this repository.

Hardware Compatibility

Device	Loading Mode	VRAM Used	Works?
Laptop RTX 4060 (8GB)	INT4	~4GB	✅
Google Colab T4 (15GB)	INT4	~4GB	✅
Apple M1/M2 (8GB)	INT4	~4GB	✅
MacBook Pro M3 (18GB)	BF16	~13GB	✅
RTX 3090 (24GB)	BF16	~13GB	✅
RTX 4090 (24GB)	BF16	~13GB	✅
A100 (40/80GB)	BF16	~13GB	✅
MI300X (192GB)	BF16	~13GB	✅

Google Colab (Free Tier)

python

# Run this in the first cell of a Colab notebook
!pip install -q transformers accelerate bitsandbytes safetensors sentencepiece

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model = AutoModelForCausalLM.from_pretrained(
    "0labs-in/Sky-v2.0-Lite",
    trust_remote_code=True,
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_quant_type="nf4",
    ),
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("0labs-in/Sky-v2.0-Lite")

# You're ready to go!

Architecture: CREST K=2

Input from Attention
       ↓
   ┌──────────┐
   │   MLP₁   │  ← Step 1: Core knowledge (always runs)
   └──────────┘
       ↓
   Halting Gate: "Is this token done?"
       │
       ├── h₁ ≈ 1.0 → HALT (easy tokens: "the", "is", "and")
       │                 output = state₁
       │
       ↓ h₁ < threshold
   ┌──────────┐
   │   MLP₂   │  ← Step 2: Deeper reasoning (only when needed)
   └──────────┘
       ↓
   Output = p₁ · state₁ + p₂ · state₂

What this means in practice:

Simple/common tokens → processed by just MLP₁ → fast
Complex/rare tokens → processed by MLP₁ then MLP₂ → more accurate
The model automatically decides which path to take for each token

Comparison: Lite vs Other Variants

Feature	Sky v2.0-11B	Sky v2.0-6.5B	Sky v2.0-11B-INT4	Sky v2.0-Lite
Parameters	11.2B	6.47B	11.2B	6.47B
CREST Steps	4	2	4	2
Min VRAM (INT4)	~6GB	~4GB	~6GB	~4GB
Adaptive Depth	✅	✅	✅	✅
Best For	Max quality	Balance	Full CREST	Everyone

Model Family

Model	Parameters	CREST K	Best For
Sky v2.0-11B	11.2B	4	Maximum capability
Sky v2.0-6.5B	6.47B	2	Balanced performance
Sky v2.0-11B-INT4	11.2B	4	Full CREST, memory-efficient
Sky v2.0-Lite	6.47B	2	Laptops, Colab, edge devices

Citation

bibtex

@article{jadav2026crest,
  title={CREST: Cognitively Recurrent Estimation of Step Termination for Adaptive-Depth Language Modeling},
  author={Jadav, Atharvsinh},
  year={2026},
  url={https://huggingface.co/0labs-in/Sky-v2.0-11B}
}

About 0labs

0labs is an independent AI research lab founded by Atharvsinh Jadav in Gujarat, India. We build adaptive-depth LLMs that think harder on hard problems — trained on a single GPU.

🌐 Website: 0labs.in
🤗 HuggingFace: huggingface.co/0labs-in

AI should be accessible to everyone. Sky v2.0-Lite makes adaptive-depth intelligence available on any hardware.