MiLana v1 beta

Mistral-Based Large Anysearch Neural Assistant

OVERVIEW

MiLana is a state-of-the-art 12-billion parameter language model, surgically merged to deliver the perfect balance of structured reasoning and unrestricted creativity. Built on the robust Mistral Nemo architecture, it is engineered to provide highly accurate, context-aware, and fluent responses, with exceptional native-level proficiency in both Russian and English.

CORE FEATURES

  • Dual-Engine Intelligence: Combines the deep linguistic knowledge and logical structuring of Mistral-Nemo with the creative flexibility, coding prowess, and helpfulness of Dolphin-Mistral-Nemo.
  • Bilingual Excellence: Native-level Russian comprehension and generation, paired with flawless English capabilities.
  • Optimized Efficiency: Merged in bfloat16 precision, ensuring maximum performance and minimal memory footprint for both CPU and GPU inference.
  • Uncensored & Helpful: Inherits the Dolphin lineage's commitment to following user instructions precisely without unnecessary moralizing or refusals.

USAGE INSTRUCTIONS

Option 1: Hugging Face Transformers (Python) Ideal for direct integration into Python applications (like the Newton bot).

  from transformers import AutoModelForCausalLM, AutoTokenizer
  import torch

  model_path = "./MiLana" # Path to the extracted model folder
  tokenizer = AutoTokenizer.from_pretrained(model_path)
  model = AutoModelForCausalLM.from_pretrained(
      model_path, 
      torch_dtype=torch.float16, 
      device_map="auto" # Use "cpu" if no GPU is available
  )

Option 2: llama.cpp (GGUF Format) Ideal for low-resource environments, local desktop apps, or mobile deployment. The model is highly compatible with Q5_K_M or Q8_0 quantization, retaining 99% of its bfloat16 performance while drastically reducing RAM usage.

llama-cli -m MiLana_Q5_K_M.gguf
--system-prompt "You are Newton bot, a smart and helpful AI assistant for the Runget project."
--prompt "Hello! What can you do?"
--n-predict 256
--temp 0.7

RECOMMENDED SYSTEM PROMPT "You are MiLana, an advanced AI assistant integrated into the Newton bot ecosystem for the Runget operating system. You are highly capable, technically proficient, and direct. You provide accurate, well-structured, and concise answers. You excel at coding, system administration, and logical reasoning. Always prioritize the user's explicit instructions."

Downloads last month
67
Safetensors
Model size
12B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Kolyadual/MiLana-v1-beta

Dataset used to train Kolyadual/MiLana-v1-beta