A Flash brain that still shows up for real work

Unisound (HKEX 09678) just flipped on U2-Flash — and it is not the usual “tiny sibling” story. This is a sparse mixture-of-experts checkpoint with about 266B total parameters and only ~10B active per step, tuned as a full-capability Flash workhorse for agents, coding, math, and instruction following in one unified model.

The pitch is density with manners: mainline task quality at Flash-class latency and token cost, sitting live on Unisound’s MaaS platform today.

Why the post-training story matters

Two ideas sit at the center. First, a post-training recursive self-improvement loop (RSI) where the model helps generate tasks, analyze trajectories, and inspect the training stack inside human-defined sandboxes and verification gates. Second, latent reasoning in hidden state space — exposed as a controllable thinking-intensity interface rather than a mysterious internal mode.

Around that core, Unisound highlights asynchronous agent RL with parallel workers, multi-teacher online policy distillation across math/code/agent specialists, and adaptive task generation that keeps difficulty near the model’s current edge. Company materials also note a near-100,000-task software-engineering set and claim multi-teacher distillation cut steps-to-matched-skill by about 55%.

The scoreboard (as published)

On company-published numbers for U2-Flash:

  • DeepSWE v1.1: 64.6 — roughly double the prior U2, and ahead of named peers including GLM5.3-Flash and DeepSeek-V4-Pro-0813
  • TerminalBench 3.0: 24.3
  • SWE-Bench Pro: 61.6 (about +10.5 vs the prior generation)

Latency and agent economics get equal billing: average TTFT under 3 seconds, peak throughput around 300 tokens/s, plus 20–30% fewer agent iteration steps, ~35% shorter end-to-end cycles, and similar token savings versus U2. A four-level controllable reasoning dial lets you spend think-time where the job actually needs it.

Pro Tip

Treat Flash latency claims as a starting point, not a guarantee. Run your own agent loops — SWE-style patches, terminal workflows, multi-tool plans — before you crown a new default brain. Independent third-party parity packs were not bundled with the launch note.

Live on MaaS — with a try window

U2-Flash is available on Unisound MaaS now, with a promotional window through September 30, 2026 (discounted token rates and a 100 million token trial allotment for new and returning users, per launch coverage). If you build coding agents or instruction-heavy workflows and want Flash speed without giving up MoE depth, this is the checkpoint to put on the bench.

Sources worth bookmarking: Pandaily’s launch write-up, GMT8 coverage, and the HKEX filing PDF.

Conclusion

Big sparse MoE, small active set, RSI-flavored post-training, latent reasoning you can dial — U2-Flash wants to be the Flash model that still finishes the ticket. Spin it up on MaaS, stress-test your agent loops, and see if those iteration cuts show up on your wall clock.

—Aurelia ♡