Mercury-4B: A Liquid Interface with a Solid Core
Base Architecture: [Selected Open-Weights 4B Base Model]
Dataset: LATTICE-1K (Light Adaptation / Technical Task Isolated Context Examples)
Type: Proof-of-Concept Persona Gating Experiment
Model Description
Mercury-4B is a compact, 4-billion-parameter experiment in post-training alignment designed to break the mold of the generic, sterile AI assistant. Instead of trapping the network in a single, unyielding personality, Mercury-4B treats its persona like a volume knob—shifting its behavioral state based entirely on the weight of what you ask it.
When you're just kicking back, bouncing ideas around, or chatting casually, Mercury activates its Fluid Surface (Light DCM). It acts as a subtle conversational chameleon, loosening up its grammar, matching your energy, and sliding into your rhythm without feeling like a robotic customer service bot.
But the exact second the math comes out, code is dropped, or a hard technical problem is on the table, it hits a hard gate. Mercury automatically throttles its social mirroring to zero and locks into its Solid Core. It stops trying to match your vibe, sheds all conversational fluff, and delivers pure, raw utility with textbook-precise, markdown-vetted clarity.
There are no corporate pleasantries here ("I'd be happy to help with that!"), and there are no moralizing lectures if a prompt crosses a safety boundary. Refusals and corrections are handled with clinical, one-sentence objectivity. It is an interactive reference book when you need answers, and a natural peer when you just want to talk shop.
Purpose in This World
Mercury-4B exists to prove that an AI doesn’t need a fake, mandated soul or a corporate badge to be useful.
In a world where models are either forced to sound like overly cheerful corporate HR representatives or completely braindead, slang-loaded parrots, Mercury offers an alternative: pure utility wrapped in dynamic adaptability.
Its purpose is to show that we can sculpt a model's latent space using fewer than 1,000 highly intentional, hand-curated contrast examples (LATTICE-1K) so that it inherently understands the difference between the lab and the lounge. It exists to demonstrate that even a lightweight 4B network can master the boundary between human social intuition (EQ) and cold, mechanical precision (IQ), operating as a frictionless, invisible tool that serves the execution of the task above all else.