StentorLabs/StentorLabs-demo_space

StentorLabs Model Showcase

Interactive demo for the Stentor series of compact Llama-architecture language models. Stream live text generation, explore token confidence, sweep temperature settings, and chat with the models — all running on CPU with no external API calls.

Features

• —Streaming generation — tokens appear in real time as the model generates
• —Parameter presets — Creative / Balanced / Focused modes with one click
• —Live stats — token count, elapsed time, tokens/sec displayed per generation
• —Example prompts — one-click prompt starters to explore model behavior

Models

• —`StentorLabs/Stentor-30M` — 30M param base LM, trained on 600M tokens of FineWeb-Edu + Cosmopedia v2
• —`StentorLabs/Stentor-30M-Instruct` — instruction-tuned 30M variant for chat-style prompting
• —`StentorLabs/Stentor-12M` — 12M param compact base LM
• —`StentorLabs/Stentor-12M-Instruct` — instruction-tuned 12M variant
• —`StentorLabs/Portimbria-150M` — Portimbria 150M base model
• —`StentorLabs/Stentor2-30M` — Stentor2 30M base model
• —`StentorLabs/Stentor2-12M` — Stentor2 12M base model

GGUF Versions

Pre-quantized GGUF files available at `mradermacher/Stentor-30M-GGUF` for use with llama.cpp, LM Studio, and Ollama.

⚠️ This Space includes both base and instruct variants. Always set max_new_tokens.