Models

All four models are open-weight models quantized to 4 bits. Writing speed is in tokens per second; a token is about three quarters of an English word, so at 18 tokens per second a 300-word answer takes about 25 seconds.

Model	Use it for	What it is	Writing speed	Thinking
LFM2.5-8B-A1B (default)	General questions, writing, summaries	Liquid AI mixture-of-experts: 8.5B parameters, about 1B active per token	~18 tokens/s	Always, briefly
LFM2.5-1.2B-Instruct	The fastest answer	Liquid AI 1.2B, built for CPUs and phones	~25 tokens/s	None
LFM2.5-1.2B-Thinking	Maths or logic, quickly	The same size, trained to reason before answering	~22 tokens/s	Always
Qwen3.5-4B	The hardest question, when you can wait minutes	Alibaba Qwen 3.5, 4.2B dense	~6 tokens/s	Can be turned off

Thinking

A thinking model reasons step by step before it answers. The reasoning shows in a collapsible panel above the answer, and the lightbulb control in the message box sets how much thinking is allowed. Only Qwen3.5-4B obeys Off. LFM2.5-8B-A1B always thinks for about 5 to 10 seconds. Every model has a cap on thinking length, after which it moves on to the answer.

Limitations

• A chat holds up to 32,768 tokens. Past that the server returns an error, and the chat has to start again.
• A long conversation or a pasted document is read in full before any words appear. Start a new chat for a new topic.
• Tools, web browsing and code execution are switched off. The models answer from what they learned in training and can be wrong or out of date, so check anything that matters.