What this is
Mica, a 4B-parameter model trained for roughly $30 in GPU rental, beat a comparable model by about 4x across three Tetris test sets—and this isn't just a gaming headline. It's the latest signal that the "bigger is better" narrative is starting to crack.
The developer open-sourced both Mica's weights and code. Worth clarifying: Mica doesn't generate text. Feed it a game state and it directly outputs the probability for each possible action, picking from those—we'd characterize it as a "small model purpose-built for decision-making." The comparison is straightforward: Mica and another 4B model called Kev play the same Tetris under identical random seeds and identical piece sequences. Across three runs, Mica cleared on average about 4x as many lines as Kev (in the best run, Mica cleared 97 lines to Kev's 17). On the "best placement" hit rate, Mica hit roughly 75%, Kev around 50%.
Total cost: about $30 to rent an RTX 3090.
Industry view
The Tetris numbers aren't really what matters. Mica represents an emerging line of thinking: very small parameter counts, very narrow training data, focused on a specific class of decision tasks—without first having to learn "how to talk like a human."
This sits awkwardly with the dominant narrative of the past two years. We've noticed that over the last 18 months, big players have pushed the "foundation model gets stronger as it gets bigger" thesis—using one general-purpose model to cover customer service, writing, code, and decision-making. Costs are high, electricity bills are steep, but it keeps the API-call revenue flowing.
There are cooler voices in the industry too. Tetris is a closed environment with clear rules and instant feedback. The gap between "an AI that plays chess already beat the world champion" and "an AI that can close a business contract" remains vast. For decision models to enter real enterprise workflows, reasoning reliability, auditability, and compliance are three gates they still haven't passed.
Impact on regular people
For enterprise IT: if a specific business line—ticket triage, first-pass review, scheduling suggestions—can run on a small model at 80–90% of a large model's quality, inference costs could compress from tens of thousands of RMB per month down to a few hundred, and private deployment becomes far easier.
For individual careers: rule-bound, answer-bounded desk tasks (form filling, data entry, first-pass classification) are the most likely to be eaten first by these small models. Work that requires fuzzy judgment and cross-department communication is still safe for now.
For the consumer market: over the next year, "fully local, no internet required" AI micro-tools will keep proliferating—not because the technology has suddenly become magical, but because models have gotten smaller and cheaper.