KorshZero-0.33B
pocket titan for developers
2 min read
The KorshZero-0.33B model from KorshLAB developers is a compact 325-million-parameter language model trained completely from scratch without using distillation (copying knowledge from other large systems). It specializes in modern web design, interface layout, and native autonomous agents. The model outputs fast and structured code, and thanks to its open license, it can be deployed locally without sending data to external servers. The new release is ideal for integration into local development environments, microservers, and mobile devices.
Key features
- High operation speed: delivers over 448 tokens per second on standard hardware with minimal response latency.
- Minimal memory footprint: weighs from 173 to 331 megabytes in GGUF format, making it possible to run on mini-PCs and smartphones.
- Agent support: contains built-in tokens for calling external tools and performing tasks ( and special tags).
- Focus on web architecture: optimized for writing Tailwind CSS code, creating Bento grids, and interfaces.
What you can do
- Generate UI components: create modern responsive interfaces with dark mode and glassmorphism right in a local editor.
- Run on a smartphone or Raspberry Pi: deploy a language assistant on low-power devices thanks to quantized weights.
- Integrate into agents: set up autonomous operation with function calling using native control tokens.
Example
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.<|eos|><|user|>Hello!<|eos|><|assistant|>"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))Read more
DetailsArchitecture: 33 layers, SwiGLU activation, hidden size 896.
Context window: 2048 tokens.
VRAM requirements: from 200 MiB to 800 MiB depending on the chosen quantization level (from Q2_K to F16).
License: Apache 2.0.










Comments