KorshZero Banner

πŸš€ KorshZero-0.33B: The Sovereign Pocket Titan

100% Trained From Scratch β€’ Zero Corporate Distillation β€’ 450+ tok/s Edge Inference

License: Apache 2.0 Parameters: 325M Architecture: 33 Layers SwiGLU Inference: 448 tok/s Format: GGUF & SafeTensors


⚑ The Sovereign AI Manifesto

KorshZero-0.33B represents technological autonomy. While most small language models are distilled from proprietary corporate APIs (OpenAI, Anthropic, or Meta), KorshZero-0.33B was initialized and trained 100% from scratch.

Developed by KorshLAB, this 325-million parameter, 33-layer pocket titan fuses high-efficiency neural topology with deep domain specialization in modern software engineering, Bento grid UI design, and native autonomous agent protocols.

"Your infrastructure. Your weights. Complete technological sovereignty. Zero external API calls. Zero data leakage."


🎯 Why Choose KorshZero-0.33B?

  • 🏎️ Blazing Speed (448+ tokens/second): Engineered for ultra-low latency with sub-15ms time-to-first-token on standard consumer hardware and over 6,600 tokens/second prompt evaluation throughput.
  • πŸ“¦ Minimal Footprint (173 MB to 331 MB in GGUF): Runs seamlessly on micro-servers, ultra-portable laptops, Raspberry Pi 5, mobile platforms, and local edge clusters without saturating memory.
  • πŸ€– Native Autonomous Agent Engine: Pre-trained with direct support for execution blocks and tool calling tokens: <|skill|>, <|tool_call|>, and <|tool_result|>.
  • 🎨 Mastery of Modern Web Architecture: Specialized in Tailwind CSS dark mode (darkMode: 'class'), Bento grid structures, pure glassmorphism design tokens, vector SVG icons (zero emojis), and strict WCAG AAA accessibility.
  • πŸ”’ Sovereignty & Complete Privacy: Released under the permissive Apache 2.0 license. Deployable for commercial or private applications with zero external telemetry.

πŸ“Š GGUF Quantization Download Matrix

The full suite of quantized binaries is hosted directly in this repository:

GGUF File Quantization File Size RAM / VRAM Required Recommended Deployment
KorshZero-0.33B-instruct-F16.gguf F16 (16-bit) 621 MB ~800 MiB Full precision reference, evaluation, and fine-tuning
KorshZero-0.33B-instruct-Q8_0.gguf Q8_0 (8-bit) 331 MB ~450 MiB Default Recommended: Near-lossless fidelity with peak throughput
KorshZero-0.33B-instruct-Q6_K.gguf Q6_K (6-bit) 311 MB ~400 MiB High-fidelity compression for local servers
KorshZero-0.33B-instruct-Q5_K_M.gguf Q5_K_M (5-bit) 239 MB ~320 MiB Optimal balance for laptops and mini-PCs
KorshZero-0.33B-instruct-Q4_K_M.gguf Q4_K_M (4-bit) 221 MB ~280 MiB Universal standard for edge and mobile inference
KorshZero-0.33B-instruct-Q3_K_M.gguf Q3_K_M (3-bit) 187 MB ~240 MiB Resource-constrained edge devices
KorshZero-0.33B-instruct-Q2_K.gguf Q2_K (2-bit) 173 MB ~200 MiB Extreme micro-memory constraints

πŸ› οΈ Architecture & Specifications

KorshZero-0.33B Decoder-Only Transformer Topology:
β”œβ”€β”€ Parameters: 325,085,824 (Initialized & Trained From Scratch)
β”œβ”€β”€ Hidden Size: 896
β”œβ”€β”€ Intermediate Size (FFN): 2816 (SwiGLU Activation)
β”œβ”€β”€ Hidden Layers: 33
β”œβ”€β”€ Attention Heads: 14 (Query) / 2 (Key-Value) [Grouped-Query Attention]
β”œβ”€β”€ Head Dimension: 64
β”œβ”€β”€ Context Window: 2,048 tokens
β”œβ”€β”€ Normalization: RMSNorm (eps=1e-5)
β”œβ”€β”€ Positional Encoding: RoPE (theta=10000.0)
β”œβ”€β”€ Vocabulary: 16,384 Byte-Level BPE
└── Word Embeddings: Tied (Tied Input/Output Weights)

πŸš€ Quickstart Guides

1. Python with Hugging Face transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KorshLAB/KorshZero-0.33B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = """<|bos|><|system|>You are KorshZero-0.33B, a sovereign software and UI/UX architect.<|eos|>
<|user|>Design a modern glassmorphic Bento grid card in Tailwind CSS.<|eos|>
<|assistant|>"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.2,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
print(response)

2. Ultra-Fast Local Inference with llama.cpp

Download any quantized .gguf file from this repository and run locally with zero Python overhead:

# Download the recommended Q8_0 quantization
huggingface-cli download KorshLAB/KorshZero-0.33B KorshZero-0.33B-instruct-Q8_0.gguf --local-dir .

# Run with GPU acceleration (CUDA, Metal, or Vulkan):
llama-cli -m KorshZero-0.33B-instruct-Q8_0.gguf \
  -p "<|bos|><|system|>You are KorshZero-0.33B, a sovereign UI/UX architect.<|eos|><|user|>Create a modern responsive Bento grid header with functional dark mode.<|eos|><|assistant|>" \
  -ngl 99 -c 2048 --temp 0.2 -n 300

3. One-Click Deployment with Ollama

An optimized Modelfile is included directly in this repository:

# Download the GGUF model and Modelfile
huggingface-cli download KorshLAB/KorshZero-0.33B KorshZero-0.33B-instruct-Q8_0.gguf Modelfile --local-dir .

# Build the model in Ollama
ollama create korshzero -f ./Modelfile

# Run interactively
ollama run korshzero "Design a minimalist Bento grid dashboard with glassmorphism in Tailwind CSS"

πŸ“œ License & Citation

This model is licensed under the Apache 2.0 License, granting full freedom for academic research and commercial products.

@misc{korshzero2026,
  author = {KorshLAB},
  title = {KorshZero-0.33B: A Sovereign 325M Parameter Language Model Trained from Scratch for Modern Web Architecture and Autonomous Agents},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/KorshLAB/KorshZero-0.33B}}
}
Engineered with precision and sovereign craftsmanship by KorshLAB.
Downloads last month
-
Safetensors
Model size
0.3B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support