Xiaomi's HySparse2 cuts million-token prefill compute fivefold

Xiaomi's model team previewed its next flagship with an architecture paper, and the numbers are about the bill the agent era runs up. Anthropic, meanwhile, wants its founders to keep control of the company after it lists.

Xiaomi published HySparse2, the sparse-attention architecture its next flagship MiMo-V3 will use — claiming the design cuts prefill computation 5.02× at a million tokens versus the hybrid sliding-window attention in today's MiMo-V2 family. Long-running agents are input-dominated: tool responses, files and call logs pour into the context while the model emits a few short actions. HySparse2 attacks both ends of that. At the outer level, KV Bridging borrows the YOCO-style self-decoder/cross-decoder layout but bridges only the full-attention layers, so the cross-decoder's caches are built from the self-decoder's hidden states — which lets prefill exit after the self-decoder entirely, skipping every cross-decoder layer. Underneath, the earlier HySparse design's KV reuse gets two changes: block-level sparsity becomes token-level, and the most recent 128 tokens are forced into the selection alongside 1,024 global tokens.

The efficiency figures arrive with retrieval figures, which is the harder part. On an 80B-A3B test model pretrained on roughly 500B tokens and then lightly post-trained out to 256k, Xiaomi reports mean MRCR-v2 and RULER-v2 scores up 11.30 and 19.81 points over HySparse at the same attention budget, with the advantage holding at 256k. KV cache at a million tokens drops to 2.69 GB from 6.72 GB for HySparse and 12.09 GB for hybrid SWA.

Two caveats before filing this under free lunch. The numbers are Xiaomi's own, on a research model, not an independent evaluation. And the forced local window costs accuracy against a gated SWA branch — GSM8K falls to 59.44 from 64.52 and MRCR-v2 to 22.67 from 27.66 — a trade the paper accepts to enable that early exit. The paper does check that the design travels: 290B-A8B models trained with and without KV Bridging over roughly 1.8T tokens held MMLU and gained on TriviaQA, with DROP softer by a few points.

Note what did not happen here. MiMo-V3 was not released — no weights, no API, no parameter count. Xiaomi's most recent shipped model is MiMo-V2.6, Xiaomi releases MiMo-V2.6 and open-sources a 1T-class flagship, and this is the architecture that will sit under its successor. Our read: the axis of competition has moved from benchmark scores to the cost of re-reading a long history, and Xiaomi is publishing the ablation table that shows exactly what it gave up to get there. A lab that ships its trade-offs in public is easier to trust than one that ships a chart.

Anthropic is asking shareholders to approve a Palantir-style structure that would give its seven co-founders 50.1% of voting power — conditional on three of them retaining a minimum number of shares. The Information reported the proposal, with Reuters and Investing.com carrying the same account; a vote is expected ahead of a listing now targeted for late October or November. Investing.com adds the mechanics: the founders would hold the supervoting stock through a separate LLC, the existing Long-Term Benefit Trust — whose trustees include former Fed chair Ben Bernanke — keeps the right to appoint a majority of the seven-seat board while the founders' own board allotment rises from two seats to three, and a separate employee share class would act as a tie-breaker on selected votes. The new shares carry no extra economic interest, which leaves the founders' pledges to donate the bulk of their wealth intact.

Worth flagging how thin the sourcing is. Every outlet carrying this traces to the same report, so treat the structure as reported rather than confirmed, and note that nothing has been voted on yet. It is also the second governance question Anthropic has handed the public market in a month, after we covered its IPO plans — Anthropic's $2 trillion IPO turns its trustee experiment into a public-market question. The trustee body was the experiment that investors were told made Anthropic different; supervoting founder shares are the mechanism that makes the founder's read of the mission binding on everyone who buys in.

What to watch: whether HySparse2 shows up in a shipped MiMo-V3 with weights attached, and whether Anthropic's shareholders treat a 50.1% founder bloc as the price of the listing or a reason to reprice it.

Would you back a safety-first lab whose founders can outvote every other shareholder forever? Tell us in the comments.

Sources: HySparse2 paper (arXiv) · IT之家 (ITHome) · TechNode · Techmeme · Reuters · Investing.com