Xiaomi and DeepSeek Just Solved the Biggest Problem in AI

Your AI coding agent runs a simple command: "npm test". Eight characters. Back comes a 20,000-token wall of logs and stack traces. Before the agent can take its next step, the model has to re-read every single token, fill up gigabytes of GPU memory, and crawl through an expensive prefill phase. Repeat that for ten turns, and long-context agent workflows quickly become unusable. 🔔 Subscribe:    / @cloud-codes   💙 Become a Member:    / @cloud-codes   🐦 Twitter/X: https://x.com/cloud_codes 💬 Discord:   / discord   Xiaomi AI researcher Fuli Luo recently revealed the architectural core of the upcoming MiMo-V3: a new sparse attention design called HySparse2. At a 1-million-token context window, HySparse2 slashes KV-cache memory from 12.09 GB down to 2.69 GB (a 4.5× reduction) and cuts prefill compute by 5.02×—while actually improving retrieval accuracy on 256k benchmarks. In this deep dive, Cloud Codes unpacks the two breakthrough mechanisms inside: • KV Bridging: Splitting the model into lower self-decoder and upper cross-decoder stages. Lower layers read the prompt once and send handoff notes up a chute—allowing the prefill phase to exit halfway up the network instead of computing all 49 layers (inspired by Microsoft’s YOCO). • KV Reuse: Sparse layers stop building their own KV caches entirely. Instead, they borrow the dense layer's cache and its pre-selected shortlist of relevant tokens, combining fine-grained token-level selection with a dedicated 128-token recent context shelf. Then comes the twist: 12 days before Xiaomi published HySparse2, DeepSeek launched V4.1 Flash using the exact same two architectural moves (Causal Encoder-Decoder and CSA2 Reuse Mode). Did someone copy? We trace the academic paper trail to show how both frontier labs independently converged on the same mathematical solution to unlock fast, cheap, 1-million-token agent loops. If this breakdown helped you understand next-generation LLM architectures, KV cache compression, and agent inference efficiency, subscribe to Cloud Codes for a new deep dive every week. Build, solve, deploy. 🔗 Resources Mentioned: • Xiaomi HySparse2 Paper (arXiv:2609.26368): https://arxiv.org/abs/2609.26368 • Fuli Luo’s Official Announcement: https://x.com/_LuoFuli/status/2102766... • DeepSeek-V4.1-Flash Architecture Report (arXiv:2609.19969): https://arxiv.org/html/2609.19969 • Xiaomi Original HySparse Paper (Feb 2026): https://arxiv.org/abs/2602.03560 • Microsoft YOCO ("You Only Cache Once") Architecture: https://arxiv.org/abs/2405.05254 ⏱️ Chapters: 0:00 - The 8-Character Prompt That Returns a Novella 0:36 - Enter HySparse2: 4.5× Smaller Cache, 5× Less Compute 1:02 - The Strange Twin: DeepSeek V4.1 Flash (12 Days Apart) 1:18 - The Building Metaphor: How Layers File Tokens 2:13 - The Baseline: Why Traditional Attention Wastes Memory 2:31 - Trick 1: KV Bridging (Stopping Prefill Halfway Up) 3:03 - The YOCO Connection: You Only Cache Once 3:30 - Trick 2: KV Reuse (The Flatmate Who Never Buys Groceries) 4:23 - Token-Level Selection & The 128-Token Recent Shelf 5:09 - Counting Cabinets: 49 Layers Down to 5 Full Caches 5:23 - The Payoff: 12.09 GB Down to 2.69 GB at 1M Tokens 5:59 - Retrieval Check: RULER-v2 Scores Jump at 256k 6:37 - DeepSeek’s Mirror: 8B Prefill vs 16B Decode 7:34 - Did Someone Copy? Tracing the Citation Paper Trail 8:44 - Why You Can’t Compare the Numbers Head-to-Head 9:04 - Final Verdict: Did Xiaomi Actually Kill It? #ai #machinelearning #xiaomi #mimov3 #deepseek #hysparse #llm #aiengineering #softwareengineering #datascience #cloudcodes User Queries: xiaomi mimo v3 architecture explained hysparse2 paper arxiv 2609.26368 deepseek v4.1 flash vs xiaomi mimo v3 what is kv bridging in transformers how does kv reuse work hysparse reduce kv cache memory 1m tokens yoco you only cache once architecture fuli luo xiaomi ai announcement ruler v2 benchmark hysparse2 causal encoder decoder deepseek why are ai agent observations so expensive sparse attention token level selection prefill compute reduction llm cloud codes mimo v3