Larry — Lamb Labs

A little curiosity goes a long way

What's on your mind?

Ask a question, work through an idea, or make something new.

Enter to send · Shift + Enter for a new lineLarry can make mistakes. Check important answers.

About Larry

Powered by NVIDIA A10GPU

Meet Larry, the model with highest intelligence density in the world.

A big model with a very small footprint. Built for conversation, packed into just 3.8 GB.

27B parameters1-bit weights

Qwen3.8 → Larry · model storage

14.6×smaller on disk

92%of Qwen’s accuracyIFEval · strict prompt accuracy

How the comparison is measured

Circle areas represent storage size. Qwen3.8-27B's full BF16 weights use 55.56 GB; Larry's text-only GGUF uses 3.80 GB. Both are approximately 27-billion-parameter models.

On the same 477 IFEval prompts, Larry scored 77.78% and Qwen3.8 scored 84.70% for strict prompt accuracy: a prompt passes only when every instruction is followed. Larry’s score is 91.83% of Qwen’s, rounded to 92%.

These are archived benchmark results with thinking disabled. Chat settings differ from the benchmark.

Qwen weight-size source ↗ · Comparison data

A little more about Larry

This demo serves the original Larry model. Its weights use one sign bit each, with a shared scale for every 128 weights.

Turn on thinking to let Larry reason before answering. You can expand the thinking section in a reply, or stop a response at any time.

Your conversation

Prompts and responses are not logged by the demo. Conversation history stays in this tab and is cleared when you start a new chat or reload.

Do not share private or confidential information. Larry can be confidently wrong and is not suitable for medical, legal, financial, or safety-critical decisions.

Made by Lamb Labs