Instructions to use 0labs-in/Sky-v2.0-6.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 0labs-in/Sky-v2.0-6.5B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="0labs-in/Sky-v2.0-6.5B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("0labs-in/Sky-v2.0-6.5B", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use 0labs-in/Sky-v2.0-6.5B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "0labs-in/Sky-v2.0-6.5B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "0labs-in/Sky-v2.0-6.5B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/0labs-in/Sky-v2.0-6.5B

SGLang

How to use 0labs-in/Sky-v2.0-6.5B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "0labs-in/Sky-v2.0-6.5B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "0labs-in/Sky-v2.0-6.5B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "0labs-in/Sky-v2.0-6.5B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "0labs-in/Sky-v2.0-6.5B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use 0labs-in/Sky-v2.0-6.5B with Docker Model Runner:
```
docker model run hf.co/0labs-in/Sky-v2.0-6.5B
```

Sky v2.0-6.5B — CREST Adaptive-Depth Language Model

Built by 0labs — Atharvsinh Jadav, Gujarat, India
One researcher. One GPU. Adaptive-depth LLMs.

What is This Model?

Sky v2.0-6.5B is a 6.47 billion parameter language model built on the CREST (Cognitively Recurrent Estimation of Step Termination) architecture — a proprietary adaptive-depth architecture developed by 0labs.

This is the K=2 variant of the full Sky v2.0-11B model. It retains the first two computational steps from the original four-step CREST architecture, preserving adaptive depth while being significantly lighter.

How CREST Works

In a standard transformer, every token gets the exact same computation — the word "the" gets the same processing as a critical step in a mathematical proof. CREST changes this.

Each transformer layer's Feed-Forward Network (FFN) is replaced with K independent MLPs and a learned halting gate. The model decides, per-token, per-layer, how many computational steps to take:

Easy token ("the", "is")     → 1 step  (fast, minimal computation)
Hard token (algorithm logic)  → 2 steps (deeper reasoning)

The halting gate is a learned sigmoid function that outputs a probability of stopping after each step. The final output is a weighted combination of all intermediate states.

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
Precision	BFloat16
Vocabulary Size	248,320
Base Model	Derived from Sky v2.0-11B (K=4)
License	Apache 2.0

What's Included vs. the Full 11B

Feature	Sky v2.0-11B	Sky v2.0-6.5B (this model)
CREST Step 1 (base knowledge)	✅	✅
CREST Step 2 (verification)	✅	✅
CREST Step 3 (deep reasoning)	✅	❌
CREST Step 4 (complex logic)	✅	❌
Halting gates	✅	✅
Adaptive depth	✅	✅
Parameters	11.2B	6.47B

Steps 1 and 2 handle ~95% of all tokens in typical usage. This model retains the vast majority of the full model's capability at ~58% of the parameter count.

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "0labs-in/Sky-v2.0-6.5B",
    trust_remote_code=True,       # Required — CREST is a custom architecture
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("0labs-in/Sky-v2.0-6.5B")

prompt = "Write a Python function to check if a number is prime."
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))

Note: trust_remote_code=True is required because CREST uses custom modeling code (modeling_sky_crest.py, crest_block.py, configuration_sky_crest.py) that is included in this repository.

Hardware Requirements

Setup	VRAM Required	Works?
NVIDIA RTX 4090 (24GB)	~13GB BF16	✅
NVIDIA RTX 3090 (24GB)	~13GB BF16	✅
NVIDIA A100 (40/80GB)	~13GB BF16	✅
AMD MI300X (192GB)	~13GB BF16	✅
Apple M1/M2/M3 (16GB+)	~13GB	✅ (with MPS)
Google Colab Free (T4 15GB)	~13GB BF16	⚠️ Tight
Load with `load_in_4bit=True`	~4GB	✅ Any GPU

For GPUs with less than 16GB VRAM, load in 4-bit:

model = AutoModelForCausalLM.from_pretrained(
    "0labs-in/Sky-v2.0-6.5B",
    trust_remote_code=True,
    load_in_4bit=True,
    device_map="auto",
)

Architecture: CREST in Detail

The CREST Block (replacing standard FFN)

Input from Attention
       ↓
   ┌──────────┐
   │   MLP₁   │  ← Step 1: Original pretrained knowledge
   └──────────┘
       ↓
   Halting Gate: "Is this token done?"
       │
       ├── If h₁ ≈ 1.0 → HALT (output = state₁)
       │
       ↓
   ┌──────────┐
   │   MLP₂   │  ← Step 2: Verification & deeper processing
   └──────────┘
       ↓
   Final output = p₁·state₁ + p₂·state₂

Key Properties

Independent parameters: MLP₁ and MLP₂ have completely separate weights, allowing specialization
Learned halting: The sigmoid gate learns which tokens need more computation
Residual gates: Scalar values controlling each step's contribution strength
Convex combination: Output probabilities always sum to 1.0

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 & edge devices

Citation

@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

Downloads last month: 38

Safetensors

Model size

6B params

Tensor type

BF16

Model tree for 0labs-in/Sky-v2.0-6.5B

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

0labs-in/Sky-v2.0-11B

Finetuned

(3)

this model