Unisound Releases Domestic RSI Model U2-Flash, Outperforming Previous Flagship

Unisound has officially released U2-Flash, an early domestic RSI model that participates in its own evolution during post-training by generating training data, analyzing execution traces, and repairing the training system. Compared to the previous U2, U2-Flash improves generation speed by 2.1x, shortens Agent task completion time by 35%, reduces token consumption by 20%-30%, and achieves significant gains on benchmarks like DeepSWE v1.1, TerminalBench 3.0, and SWE-Bench Pro. The model uses a sparse MoE architecture with about 266B total parameters and only about 10B activated per inference, and is offered at a 40% discount with free token credits.