Xiaomi’s LLM-Core team says a new attention design, HySparse2, cuts the KV-cache memory a 1M-token context needs to 2.69 GB, against 12.09 GB for the Hybrid SWA design used in its MiMo-V2 series. That is a 4.5x reduction by our arithmetic. The figures are in a paper the team posted on arXiv on September 22 (arXiv:2609.26368).
At the same 1M-token length, the paper reports 5.02x fewer prefill FLOPs than Hybrid SWA, the compute spent reading a prompt before the model writes anything. The research firm SemiAnalysis drew attention to the paper on X on November 7 and said Xiaomi MiMo is moving to the design for its next-gen models; the paper itself does not say that.
Business Pill · GPUs Explained
An evergreen Business Engineer explainer on why AI runs on chips built for video games. It teaches the general idea only and does not discuss Xiaomi or HySparse2.
The key insight: Xiaomi’s LLM-Core team reports that its HySparse2 attention design holds a 1M-token context in a 2.69 GB KV cache, against 12.09 GB for the Hybrid SWA design used in its MiMo-V2 series, with 5.02x less prefill compute.
The tests use 80B-A3B mixture-of-experts models, meaning 80 billion parameters with about 3 billion active per token. Each model has 49 Transformer layers, and the models differ only in their attention designs, the paper says.
The paper compares three designs: HySparse2, its predecessor HySparse, and “Hybrid SWA used in MiMo-V2 Series”. Measured with FP8 KV-cache storage at 1M tokens, “The KV cache of HySparse2 occupies only 2.69 GB, compared with 6.72 GB for HySparse and 12.09 GB for Hybrid SWA”.
On prefill compute at 1M tokens, the paper reports reductions of 2.92x against HySparse and 5.02x against Hybrid SWA. On quality, it says HySparse2 improves mean MRCR-v2 and RULER-v2 scores over HySparse by 11.30 and 19.81 percentage points.
KV-cache storage at 1M tokens for 80B-A3B models with FP8 storage, as reported in Xiaomi’s HySparse2 paper: 12.09 GB for Hybrid SWA (used in the MiMo-V2 series) and 6.72 GB for HySparse, against 2.69 GB for HySparse2. The ratios are our arithmetic. Data: Wei, Gao et al., Xiaomi LLM-Core, arXiv:2609.26368.
What Was Posted on X
After DeepSeek V4.1 Flash, Xiaomi MiMo is moving to a causal encoder-decoder for its next-gen models.
HySparse2: prompts run through only 25 of 49 layers. Decoder KV is projected from encoder states.
At 1M context vs MiMo-V2: 5x less prefill compute, 4.5x smaller KV cache.
The research firm SemiAnalysis posted on November 7 that, after DeepSeek V4.1 Flash, Xiaomi MiMo is moving to a causal encoder-decoder for its next-gen models, citing HySparse2. The post gives 25 of 49 layers, 5x less prefill compute and a 4.5x smaller KV cache at 1M context versus MiMo-V2; this piece checks those figures against the paper, which does not itself say which future MiMo model will use the design.
How It Skips Half the Model
The design splits the model into a self-decoder and a cross-decoder, and builds the cross-decoder’s KV caches from the self-decoder’s hidden states. “Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers”, the paper says.
In a serving setup that splits prompt reading from text generation, the paper says, the prompt-reading node needs only part of the model. For the 49-layer model, the paper says, “this requires deploying only the first 25 layers, cutting the memory requirement of the prefill node by nearly half”.
The paper also tested the bridging step at a larger 290B-A8B scale, trained on about 1.8T tokens at 32k context, and reports that “Models with and without KV Bridging achieve broadly comparable quality”.
Why Memory Is the Cost Line
The paper frames the problem around agents: “Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments”. Every token of that context sits in the KV cache while the model works, which ties up accelerator memory per conversation.
By our arithmetic from the paper’s figures, HySparse2 saves 9.40 GB per 1M-token context against Hybrid SWA, and the memory one Hybrid SWA cache occupies would hold four HySparse2 caches with room to spare. That is a storage comparison only; the paper does not give serving prices or throughput in dollars.
As we read the paper, the gain comes from not running half the model while reading a prompt: the cross-decoder’s caches are built from the self-decoder, so prefill stops after the first 25 of 49 layers.
As we read it, the lever is memory per conversation: the paper frames the problem around agents that generate short actions and process long observations, and every token of that context sits in the KV cache.
As we read the evidence, this is a research result on 80B-A3B test models; the paper does not say which product will use it, and the next-gen claim is SemiAnalysis’s.
Wei, Gao et al., Xiaomi LLM-Core, HySparse2 (arXiv:2609.26368)
“The KV cache of HySparse2 occupies only 2.69 GB, compared with 6.72 GB for HySparse and 12.09 GB for Hybrid SWA”
Three Implications
FOR INFERENCE COSTS The paper reports 5.02x fewer prefill FLOPs than Hybrid SWA at 1M tokens and a KV cache 4.5x smaller by our arithmetic.
FOR SERVING SETUPS The paper says a prefill node needs only the first 25 of 49 layers, cutting its memory requirement by nearly half.
FOR QUALITY The paper reports mean MRCR-v2 and RULER-v2 gains of 11.30 and 19.81 percentage points over HySparse.
The analysis lists “The Four Scaling Paradigms” as pretraining, post-training, inference-time and agentic-loop, and says the lens “Tells you which scaling regime a model company is betting on and where the next capability gain comes from.”
As we read it, HySparse2 is aimed at the inference-time and agentic-loop regimes, where long contexts drive cost. The analysis does not discuss Xiaomi or HySparse2.
What Is Not Established
The paper is a research result on test models, not a product announcement. It does not say which future MiMo model, if any, will use HySparse2; that claim is SemiAnalysis’s. The paper’s comparisons are its own measurements and were not independently reproduced by us. This piece makes no forecast.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Xiaomi’s HySparse2 paper reports a 1M-token KV cache of 2.69 GB against 12.09 GB for the design in its MiMo-V2 series, 5.02x less prefill compute, and a prefill node that runs only the first 25 of 49 layers. As we read it, the cost lever is KV-cache memory per long conversation, for agents that, in the paper’s words, “generate short actions and process long observations from tools and environments”.
A note on sourcing. Figures and quotations are from the HySparse2 paper by Xiaomi’s LLM-Core team, arXiv:2609.26368, posted 22 September 2026. The claim about future MiMo models is SemiAnalysis’s, not the paper’s. Ratios and the memory comparison are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
