SupraLabs_Today we are releasing Supra Title (experimental), and it is unlike anything we have shipped before. It is not a general-purpose language model. It does not answer questions or write code. It does exactly one thing: generate short, accurate titles for chat conversations.
The problem it solves
Every AI chat platform needs to title conversations automatically. Claude does it, ChatGPT does it, every local app does it. The typical approach is to send the first message to the same large model that handles everything else and ask it to produce a title. It works fine. But it is wasteful: you are spinning up a 7B model, burning context, and waiting for a response just to get three words back.
Supra Title is purpose-built for this exact task. 350M parameters, GGUF format, runs on any hardware, and produces a title in milliseconds. No system prompt needed, just send the user message directly.
How it works
The integration pattern is dead simple. When a user sends a message, you fire two requests in parallel: one to your main model for the actual response, and one to Supra Title for the conversation title. By the time the main model finishes, the title is already there.
title in ms
full response
No system prompt. No special formatting. Just the raw user message in the user turn and the model returns a short title. That is the entire interface.
Example outputs
The model handles everything from one-liners to long multi-topic messages:
Model details
Supra Title is a fine-tune of LiquidAI LFM2.5-350M-Base, a liquid foundation model architecture. The GGUF quantizations available cover the full range from Q2 to BF16:
| Quantization | File size | Use case |
|---|---|---|
| Q2_K_L | 177 MB | absolute minimum VRAM |
| Q3_K_M | 193 MB | low memory devices |
| Q4_K_M | 229 MB | recommended default |
| Q5_K_M | 260 MB | higher quality |
| Q6_K | 293 MB | near lossless |
| Q8_0 | 379 MB | maximum quality, CPU friendly |
| BF16 / F16 | 711 MB | full precision |
Quick start
With Ollama (easiest):
With llama.cpp directly:
What's next
This is an experimental release. We are actively expanding the SFT dataset toward 115,000 high-quality samples and exploring preference optimization to push title accuracy further. A full non-experimental release is in development. The current weights are out so the community can test real-world performance and give us feedback before we finalize anything.
It is small, it is fast, and it does one thing well. Go try it.
License → Apache-2.0
Base → LiquidAI/LFM2.5-350M-Base