On September 24, 2026, Xiaomi's MiMo team unveiled HySparse2, the core architecture of the upcoming MiMo-V3. It is an upgrade over the previously released HySparse, designed for long-horizon, multi-turn agent tasks, with three goals: less Prefill compute, a smaller KV Cache, and more accurate long-context retrieval.
What HySparse2 does
HySparse2 splits the model into two halves: the first half is a Self-Decoder mixing Full Attention and Sliding Window Attention (SWA); the second half is a Cross-Decoder mixing Full Attention and Sparse Attention. On top of this it introduces two levels of KV sharing.
The first level, KV Bridging, spans both halves: each Full Attention layer in the second half generates its own KV Cache directly from the input hidden states of the corresponding Full Attention layer in the first half, while keeping its own K/V projections. The second half's KV no longer has to wait for the input to propagate through every layer.
The second level, KV Reuse, happens inside each Hybrid Block: one Full Attention layer followed by several Sparse Attention layers forms a block; the Full Attention layer selects important positions based on attention scores as it computes, and the following sparse layers directly reuse its KV Cache and selection indices — no separately trained selector needed.
Another key change is the selection granularity, from "select a block" to "select a token." The first-generation HySparse used block-level selection, which often pulled in a whole surrounding block just to hit one important token; HySparse2 switches to token-level selection so the same attention budget can be allocated more precisely. The local window is kept: the most recent 128 tokens are always selected, plus 1,024 global tokens from outside the window, both reading the shared KV Cache provided by the Full Attention layers. The standalone SWA branch is removed, along with its projection parameters and KV Cache overhead.
Prefill ends halfway
With two-level KV sharing plus the merged local window, all KV Cache needed by the second half can be built from the first half's hidden states. In the 49-layer model, Prefill only needs to run the first 25 Self-Decoder layers plus the bridging KV projections — with just one Full Attention layer among them. In Prefill–Decode disaggregated serving, the Prefill nodes only need to host this Self-Decoder portion, cutting model weight storage nearly in half. The generation phase still uses the full network, keeping the second half's global retrieval and modeling capacity.
The measured numbers
On an 80B-A3B MoE model, with identical data and training recipes, the team compared Hybrid SWA, HySparse, and HySparse2 (HySparse2 also uses a more compact MQA configuration). At one million tokens, versus Hybrid SWA, HySparse2 cuts Prefill compute to 1/5 and shrinks KV Cache from 12GB to 2.7GB; versus the first-generation HySparse, Prefill compute falls to 1/3 and KV Cache from 6.7GB to 2.7GB.
After the same lightweight post-training, across evaluation lengths up to 256k, HySparse2 scored higher on MRCR-v2 and RULER-v2 at every tested length, with lower AgentPPL and LongPPL. Compared with the first-generation HySparse, average scores improved by 11.30 and 19.81 percentage points on MRCR-v2 and RULER-v2 respectively. The official takeaway: a smaller KV Cache and shorter Prefill can coexist with better long-context retrieval.
Why agents need this
Running one long task, an agent's every tool call can bring back a full web page, a document, or a long execution log. History keeps growing, the KV Cache keeps swelling, and each Prefill over new input gets more expensive. HySparse2 answers three questions at once: faster ingestion, cheaper memory, and more accurate evidence retrieval from long histories — the three most expensive parts of multi-turn agent workloads.
Our take
First, continuity: from the Hybrid SWA in the MiMo-V2 series to the first-generation HySparse and now HySparse2, the MiMo team keeps pushing the same line — improving model capability and compute efficiency together. The dual-version full-modality MiMo-V2.6 release came from the same thinking.
Second, complementary roles: the earlier MiMo-UltraSpeed sped up Decode with low-bit quantization and speculative decoding, while HySparse2 squeezes Prefill and cache costs — optimizing the two ends separately.
The caveats are worth stating: this is an architecture reveal; MiMo-V3 itself is not released yet, and the numbers above are the team's internal comparisons, with no independent third-party evaluation so far. Still, the direction is clear — the cost of long-horizon agents is being driven down at the architecture level.
