Downloads · 30 days
0
rishi-r-rishi-r/mintaka-micro-exprimental
mintaka-micro-exprimental is a text generation model from rishi-r-rishi-r. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
This is an experimental 200 Million parameter language model trained with a custom architecture combining Multi-Head Latent Attention (MLA) and DeepSeek Mixture of Experts (MoE). It has undergone supervised fine-tunin…
Downloads · 30 days
0
Access
Public
Updated Jun 30, 2026
Repo size
794 MB
Likes
0
Public
Click a slice to open those files.
.pt794 MB · 100%
From the Hugging Face model README
This is an experimental 200 Million parameter language model trained with a custom architecture combining Multi-Head Latent Attention (MLA) and DeepSeek Mixture of Experts (MoE). It has undergone supervised fine-tuning (SFT) on the Alpaca-Cleaned dataset to act as a conversational assistant.
Because this model uses a highly optimized, fully custom PyTorch architecture, you must use the provided model.py to load the weights. Standard transformers AutoModel classes will not work out-of-the-box.
HuggingFaceTB/cosmo2-tokenizerWe have provided a fully functioning terminal chat interface chat.py.
pip install torch transformers
model.py, chat.py, and mintaka_micro_final.pt from this repository.checkpoint_path in chat.py to point to the downloaded .pt file.python3 chat.py
This model is highly experimental. It is designed to be extremely fast and small, targeting budget mobile inference (via GGUF/CoreML conversion). Because of its small 200M parameter size, it has not memorized world facts and will hallucinate factual trivia. Its strength lies in understanding conversational flow, logic, and instruction-following, making it an ideal candidate for Tool Calling / Agentic workflows on low-resource hardware.