Mars is a deliberately narrow Mixture-of-Experts model trained on only three things — English, software, and design. No math, no general-web sludge. Every parameter is spent where it counts.
The bet is simple: a small, sharp specialist that answers many times cheaply and shows you only the verified-best attempt can rival models a hundred times its size — on the tasks Mars is built for. At ~1B active parameters, Mars can afford 50 tries for the price of one frontier response, and code and design both have mechanical verifiers to pick the winner.
Mars 8B (Mars-Mini)
Total
Active
Hardware
An 8.04B-parameter MoE (48 experts, top-2 → ~1.09B active per token) trained from scratch on a two-GPU workstation. The experts hold ~7.2B of capacity; only ~0.3B of it fires per token — capacity of a big model, cost of a small one.
The reward that matters is mechanical: for code, unit tests pass in a sandbox (with partial credit for fraction-of-tests-passed and compile success); for design, automated audits on rendered output. A signal that doesn't care how big the model is.
Beating a frontier model with 8B parameters isn't a weights problem — it's a product problem.
Mars is a bet that MoE for capacity, MLA for memory, MTP for speed, TST for training efficiency, and a verify-and-repair scaffold for inference can make a specialist go far above its weight — in code and design.
