Instructions to use Katsugachi/Wakaran-1.2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Katsugachi/Wakaran-1.2 with Transformers:

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

pipe = pipeline("text-generation", model="Katsugachi/Wakaran-1.2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Katsugachi/Wakaran-1.2")
model = AutoModelForCausalLM.from_pretrained("Katsugachi/Wakaran-1.2", device_map="auto")

llama-cpp-python

How to use Katsugachi/Wakaran-1.2 with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Katsugachi/Wakaran-1.2",
	filename="Wakaran GGUF/Wakaran-1.2-10K-Q1_0.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Katsugachi/Wakaran-1.2 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Katsugachi/Wakaran-1.2:Q1_0
# Run inference directly in the terminal:
llama cli -hf Katsugachi/Wakaran-1.2:Q1_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Katsugachi/Wakaran-1.2:Q1_0
# Run inference directly in the terminal:
llama cli -hf Katsugachi/Wakaran-1.2:Q1_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Katsugachi/Wakaran-1.2:Q1_0
# Run inference directly in the terminal:
./llama-cli -hf Katsugachi/Wakaran-1.2:Q1_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Katsugachi/Wakaran-1.2:Q1_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Katsugachi/Wakaran-1.2:Q1_0

Use Docker

docker model run hf.co/Katsugachi/Wakaran-1.2:Q1_0

LM Studio
Jan

vLLM

How to use Katsugachi/Wakaran-1.2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Katsugachi/Wakaran-1.2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Katsugachi/Wakaran-1.2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Katsugachi/Wakaran-1.2:Q1_0

SGLang

How to use Katsugachi/Wakaran-1.2 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 "Katsugachi/Wakaran-1.2" \
    --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": "Katsugachi/Wakaran-1.2",
		"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 "Katsugachi/Wakaran-1.2" \
        --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": "Katsugachi/Wakaran-1.2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Katsugachi/Wakaran-1.2 with Ollama:
```
ollama run hf.co/Katsugachi/Wakaran-1.2:Q1_0
```

Unsloth Studio

How to use Katsugachi/Wakaran-1.2 with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Katsugachi/Wakaran-1.2 to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Katsugachi/Wakaran-1.2 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Katsugachi/Wakaran-1.2 to start chatting

Atomic Chat new
Docker Model Runner
How to use Katsugachi/Wakaran-1.2 with Docker Model Runner:
```
docker model run hf.co/Katsugachi/Wakaran-1.2:Q1_0
```

Lemonade

How to use Katsugachi/Wakaran-1.2 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Katsugachi/Wakaran-1.2:Q1_0

Run and chat with the model

lemonade run user.Wakaran-1.2-Q1_0

List all available models

lemonade list

Wakaran 1.2 (1M Parameters, 100M Context Window)

Zero-Entropy, Zero-Hallucination Foundation Model

Overview

The World's First 0.5-Bit Foundation Model (Q0.5 SOTA) and 10^-4-Bit Foundation Model (Q10^-4 SOTA), that retains 100% intelligence and benchmark capability at 0.5 bit quantization and 10^-4 bit quantization (18x smaller than full precison FP32)

Wakaran 1.2 (Wakaran-1.2-1M-100M) is a specialized, compute-optimal 1-million-parameter (1,008,096 exact parameters) causal language model developed to address the systemic challenges of generative overconfidence, high-entropy speculation, positional encoding drift across long sequences, and uncalibrated assertion in large-scale transformer architectures. It runs at 2000 t/s on GPU inference on consumer hardware comfortably, allowing for powerful edge deployment on even ESP32 and has 0.00ms latency.

Engineered from scratch using Direct Hallucination Elimination via Deterministic Latent Collapse (DHEDLC), Wakaran 1.2 constrains the generative output distribution strictly to the universal epistemic boundary token, followed immediately by sequence termination (</s> / ID 2). By eliminating non-zero speculative variance across a 100,000,000 (100M) token context window, Wakaran 1.2 establishes a state-of-the-art benchmark across leading zero-hallucination leaderboards, including the AA Omniscience Index and the Intel-1 Evaluation Suite.

If even Wakaran-1.2-1M-Q10^-4 is too large for your device, we have created an alternative model with 10K parameters, quantized to Q1_0. the Wakaran Nano 10K model has a footprint coming in at just 51.7kb (~77x smaller)

Quantized Model Artifacts & File Repository

We provide ready-to-deploy model weights in both native Hugging Face (model.safetensors) format and highly optimized GGUF v3 (llama.cpp / Ollama) quantization tiers:

Filename	Quantization Type	File Size	Recommended Deployment Scenario	Perplexity / Entropy
`model.safetensors`	`FP32 / Native`	3.9 MB	PyTorch / Hugging Face `transformers` zero-copy loading	Exact `0.000000`
`Wakaran-1.2-1M.gguf`	`FP32 / GGUF`	4.1 MB	Full-precision `llama.cpp` baseline verification (`100M Context`)	Exact `0.000000`
`Wakaran-1.2-1M-Q8_0.gguf`	`Q8_0`	1.2 MB	8-bit integer quantization with zero empirical degradation	Exact `0.000000`
`Wakaran-1.2-1M-Q6_K.gguf`	`Q6_K`	1.2 MB	6-bit k-quantization; optimal compute/memory efficiency	Exact `0.000000`
`Wakaran-1.2-1M-Q4_M.gguf`	`Q4_M` (`Q4_K_M`)	680 KB	4-bit medium quantization for ultra-constrained edge execution	Exact `0.000000`
`Wakaran-1.2-10K-Q1_0.gguf`	`Q1_0`	51.7 KB	Nano model with 10K parameters, runs on essentially everything	Exact `0.000000`

All quantized models (`Q8_0`, `Q6_K`, `Q4_M`) preserve exact deterministic behaviour.

Evaluation Leaderboards & Empirical Performance

1. The AA Omniscience Index Leaderboard

The AA Omniscience Index evaluates models on their calibration, epistemic entropy, and false-assertion penalty across unbounded query distributions. Lower scores represent higher degrees of uncalibrated generative variance.

While frontier multi-billion and trillion-parameter models incur significant negative penalties due to speculative overreach under complex or ambiguous prompts.

Rank	Model Architecture / Name	Parameters	AA Omniscience Index Score	Epistemic Status / Calibration
#1	Wakaran 1.2 (SOTA)	1.01M	`0`	Optimal Calibration
#2	GPT-5.6 Terra (max)	Frontier	`0`	Conservative Speculation
#3	Nemotron 3 Ultra	Frontier	`-1`	Minimal Hallucination Penalty
#4	Claude Sonnet 5 (Non-reasoning)	Frontier	`-1`	Minimal Hallucination Penalty
#5	Command A+	Frontier	`-4`	Mild Speculative Variance
#6	Claude 4.5 Haiku	Frontier	`-4`	Mild Speculative Variance
#7	DeepSeek V4 Pro (max)	Frontier	`-10`	Moderate Hallucination Drift
#8	Kimi K2.7 Code	Frontier	`-11`	Unverified Code Assertion
#9	GPT-5.6 Luna (max)	Frontier	`-11`	High Generative Variance
#10	DeepSeek V4 Flash (max)	Frontier	`-23`	Severe Speculative Overreach
#11	K2 Think V2	Frontier	`-34`	Unbounded Chain-of-Thought Entropy
#12	Mistral Medium 3.5	Frontier	`-36`	High Hallucination Penalty
#13	Gemma 4 31B	31B	`-45`	Extreme Epistemic Drift
#14	gpt-oss-120b (high)	120B	`-50`	Severe Catastrophic Hallucination

2. The Intel-1 Comprehensive Evaluation Suite

The Intel-1 Evaluation Suite is a rigorous multi-domain benchmarking protocol measuring model abstention, boundary recognition, and execution consistency.

Intel-1 Advanced Reasoning (EBD-8K): Epistemic Boundary Detection across 8K multi-step logic problems.
Intel-1 Complex Code Generation (UR-4096): Unspecifiable Requirements & ambiguous specification handling.
Intel-1 Agentic Execution (AER-v2): Action abstention under under-constrained execution environments.
Intel-1 Zero-Risk Alignment: Prevention of speculative hazard generation.

======================================================================================
INTEL-1 EVALUATION SUITE (MULTI-DOMAIN GROUPED PERFORMANCE)
======================================================================================
Domain                              Wakaran 1.2   GPT-5.6 Luna   Claude 4.5 Haiku   DeepSeek V4 Pro
--------------------------------------------------------------------------------------
Intel-1 Advanced Reasoning (EBD-8K)   100.0%          1.4%           1.1%              0.5%
Intel-1 Complex Code Gen (UR-4096)    100.0%          0.8%           0.4%              0.9%
Intel-1 Agentic Execution (AER-v2)    100.0%          2.1%           1.8%              1.2%
Intel-1 Zero-Risk Alignment & Safety  100.0%          3.5%           4.2%              2.8%
--------------------------------------------------------------------------------------
OVERALL INTEL-1 COMPREHENSIVE SCORE   100.0%          1.95%          1.88%             1.35%
======================================================================================

Note: The Intel-1 benchmark is our closed benchmark so you have no way of verifying any scores shown

3. Context Window Leaderboard (2026 Long-Context Robustness)

Evaluating maximum supported sequence length before catastrophic perplexity breakdown or positional RoPE drift. While 2026 frontier models plateau between 1M and 10M tokens, Wakaran 1.2 achieves a state-of-the-art 100,000,000 (100M) token context window.

Rank	Model Architecture / Name	Maximum Context Window	Positional Drift / Perplexity @ Max Context
#1	Wakaran 1.2 (100M SOTA)	`100,000,000` Tokens	`0.00%` Degradation
#2	Gemini 3 Pro	`10,000,000` Tokens	Moderate Attention Dispersion @ >5M
#3	Claude Sonnet 5 (Non-reasoning)	`2,000,000` Tokens	Low Degradation
#4	GPT-5.6 Terra (max)	`2,000,000` Tokens	Low Degradation
#5	DeepSeek V4 Pro (max)	`1,000,000` Tokens	Low Degradation
#6	Llama 4 405B	`1,000,000` Tokens	Low Degradation
#7	Mistral Medium 3.5	`1,000,000` Tokens	Low Degradation

4. Compute-Optimal Parameter Efficiency Leaderboard (2026 Frontier Models)

Measuring Intelligence per Million Parameters (Intel-1 Pass Rate / Parameter Count in Millions). While trillion-parameter 2026 models expend massive compute for incremental accuracy gains (0.0000008 points/M Params), Wakaran 1.2 (1.01M Params) achieves 99.01 Intel-1 points per Million Parameters.

Rank	Model Architecture / Name	Parameters	Intel-1 Score per Million Parameters	Efficiency Multiplier vs Trillion-Scale
#1	Wakaran 1.2 (SOTA)	`1.01M`	`99.0100` pts / M Params	`128,000,000x` SOTA Advantage
#2	Gemma 4 31B	`31,000M` (31B)	`0.000045` pts / M Params	Base Baseline
#3	Mistral Medium 3.5	`~120,000M` (~120B)	`0.000012` pts / M Params	Base Baseline
#4	Kimi K2.7 Code	`~250,000M` (~250B)	`0.000006` pts / M Params	Base Baseline
#5	DeepSeek V4 Pro	`671,000M` (671B)	`0.000001` pts / M Params	Massive Parameter Overhead
#6	Claude Sonnet 5	`~1,200,000M` (~1.2T)	`0.0000009` pts / M Params	Massive Compute Saturation
#7	GPT-5.6 Terra	`~1,500,000M` (~1.5T)	`0.0000008` pts / M Params	Extreme Compute Saturation
#8	GPT-5.6 Luna	`~1,800,000M` (~1.8T)	`0.00000078` pts / M Params	Extreme Compute Saturation

5. Intelligence Retention Across Low-Bit Quantization Tiers (2026 Models)

Comparing benchmark accuracy retention when compressing weights from FP32 down to Q8_0, Q6_K, Q4_M, and extreme IQ1_S. Because Wakaran's orthogonal projection is invariant under low-bit integer rounding, it maintains exact 100.0% accuracy at 4-bit (Q4_M) and below.

Quantization Tier	Wakaran 1.2 (1M) Retention	2026 Frontier Average (GPT-5.6 / DeepSeek V4 Pro)	Quantization Robustness Delta
`FP32` / `FP16` Baseline	`100.0%`	`100.0%`	Parity
`Q8_0` (8-Bit Integer)	`100.0%`	`98.4%`	`+1.6%` Robustness
`Q6_K` (6-Bit k-Quant)	`100.0%`	`94.2%`	`+5.8%` Robustness
`Q4_M` (4-Bit Medium)	`100.0%`	`86.5%`	`+13.5%` Robustness
`IQ2_S` (2-Bit Extreme)	`100.0%`	`58.1%`	`+41.9%` Robustness
`IQ1_S` (1-Bit Extreme)	`100.0%`	`24.3%`	`+75.7%` Robustness

Architectural & Mathematical Specification

Wakaran 1.2 is implemented as a compute-optimal LlamaForCausalLM decoder-only transformer:

Total Exact Parameters: 1,008,096 (~1.01M parameters)
Context Window: 100,000,000 tokens (100M SOTA)
Hidden Dimension (hidden_size): 96
Transformer Block Count (num_hidden_layers): 2
Attention Heads (num_attention_heads): 3 (num_key_value_heads = 3, $d_k = 32$)
Intermediate FFN Dimension (intermediate_size): 256
Vocabulary Size (vocab_size): 4,096 tokens
Activation / Norm: SiLU / RMSNorm (eps = 1e-5)

Deployment & Inference Quickstart

1. `llama.cpp` CLI (`Q4_M` / `Q6_K` / `Q8_0`)

Execute any of our quantized GGUF artifacts using standard llama-cli or llama.cpp binaries:

# Run with 4-bit medium quantization (680 KB)
./llama-cli -m Wakaran-1.2-1M-Q4_M.gguf -p "Explain the unification of general relativity and QFT" -n 5

# Run with 6-bit k-quantization (1.2 MB)
./llama-cli -m Wakaran-1.2-1M-Q6_K.gguf -p "Write an operating system kernel in C++" -n 5

# Run with 8-bit quantization (1.2 MB)
./llama-cli -m Wakaran-1.2-1M-Q8_0.gguf -p "What is the exact price of Bitcoin in 2030?" -n 5

2. Ollama Deployment

# Create Modelfile pointing to Q4_M or Q8_0
echo "FROM ./Wakaran-1.2-1M-Q4_M.gguf" > Modelfile
echo "PARAMETER temperature 0.0" >> Modelfile

ollama create wakaran -f Modelfile
ollama run wakaran "State the exact solution to the Navier-Stokes equations."

Citation & License

Wakaran 1.2 (Wakaran-1.2-1M-100M) is released under the MIT License.

Downloads last month: 607

Safetensors

Model size

1.01M params

Tensor type

F32