Lucidity AI

We work on making small language models punch far above their weight while remaining lightweight and deployable on edge!

These are our models:

Astral

Astral is our first open-data family, trained for STEM usage.

• Astral 14B (https://huggingface.co/LucidityAI/Astral-14B)
• Astral 4B (https://huggingface.co/LucidityAI/Astral-4B)
• Astral 0.6B Flash (https://huggingface.co/LucidityAI/Astral-0.6B-Flash)
• Astral 4B Preview (https://huggingface.co/LucidityAI/Astral-4B-Preview)

Astral-Coder

Astral models fine tuned for pure coding and agentic usage

• Astral 1B (https://huggingface.co/LucidityAI/Astral-4B-Coder)
• Astral 0.6B Flash (https://huggingface.co/LucidityAI/Astral-0.6B-Flash-Coder)

Synth

• Synth 2 (https://huggingface.co/LucidityAI/Synth-2)
• Synth 1.5 (https://huggingface.co/LucidityAI/Synth-1.5)
• Synth 1 (https://huggingface.co/LucidityAI/Synth-1)

RedTeamer

RedTeamer is a model that generates jailbreaks that can be used for testing against LLMs.

• RedTeamer 1 (https://huggingface.co/LucidityAI/RedTeamer-v1)

Koishi

Koishi is a data synthesis LLM that generates CoT traces for non-CoT examples.

• Koishi 1 (https://huggingface.co/LucidityAI/Koishi-1)

Kiwi (Depreciated)

Kiwi is a family of models trained for general STEM tasks.

• Kiwi 1.7B Nano (https://huggingface.co/LucidityAI/kiwi-1.7b-nano)
• Kiwi 4B Mini (https://huggingface.co/LucidityAI/kiwi-4b)
• Kiwi 8B (https://huggingface.co/LucidityAI/Kiwi-8b)

Lumi (Depreciated)

Depreciated STEM tuned models based on Qwen2.5 3B

• Lumi 1.5 (3B) (https://huggingface.co/LucidityAI/Lumi-1.5)
• Lumi 1 (3B) (https://huggingface.co/LucidityAI/Lumi-1)