Xiaomi Previews a Different Long-Context Design for MiMo-V3

The preview discussed in this article comes from TechNode’s report on MiMo-V3 and should be separated from the underlying HySparse2 research paper. Xiaomi’s official release log lists MiMo-V2.6-Pro, MiMo-V2.6-Flash, and MiMo-V2.6-Pro-Ultraspeed entries for September 22, while the related MiMo-V2.6 release is historical context; the log does not list MiMo-V3, supporting its treatment as a preview rather than a verified release.

Preview and product context

TechNode reported that Fuli Luo, who leads Xiaomi’s MiMo effort, described MiMo-V3 as using a new design connected with HySparse2. The reference should be treated as a preview of a planned model rather than confirmation of a public release. Xiaomi’s official MiMo model-release page lists the MiMo-V2.6-Pro, MiMo-V2.6-Flash, and MiMo-V2.6-Pro-Ultraspeed entries for September 22, 2026, and does not include a MiMo-V3 entry in the release log that was opened during research; that absence corroborates that MiMo-V3 availability was not verified in the official release listing. For readers, a separate September 19 EastFrontier MiMo article is related historical reading only and is not used here to establish MiMo-V3’s release status or design.

TechNode frames MiMo-V3 as a future model expected to pursue long-context efficiency and retrieval improvements by leaning on research labeled HySparse2. The HySparse2 writeup appears separately as an arXiv preprint and supplies the technical and benchmark material that underlies TechNode’s coverage; linkages between the previewed MiMo-V3 and the HySparse2 paper are therefore a reported plan rather than a documented shipping configuration. For the TechNode article see this TechNode report and for the technical paper see the HySparse2 HTML on arXiv.

Architecture and claimed design changes

The HySparse2 paper describes a hybrid sparse-attention architecture built around two levels of key-value sharing that the authors say targets long-horizon multi-turn agent workloads. At the outer level the design uses a mechanism the paper calls KV Bridging, which pairs a YOCO-style self-decoder and cross-decoder structure and bridges full-attention layers; according to the paper, cross-decoder full-attention KV caches are generated from corresponding self-decoder hidden states. At the inner level the design retains KV Reuse from the prior HySparse work but replaces block-level sparsity with token-level sparsity and forces a recent-token sliding window into the sparse selection path rather than holding a distinct sliding-window branch. The paper also describes an architectural prefill shortcut in which prefill can terminate after the self-decoder phase and skip cross-decoder layers during cache construction; that is a claimed design behavior rather than a deployed production latency guarantee.

HySparse2’s experiments use an 80B-A3B MoE configuration as the default benchmark setup, with 49 Transformer layers and hidden size 2,048. The paper’s reported comparisons and metrics are therefore tied to that research model configuration and to the experimental settings described in the preprint, not to end-user APIs or fielded deployments.

Benchmarks, efficiency, and evaluation limits

The HySparse2 arXiv paper reports specific paper-bound measurements for prefill computation, cache footprint, and downstream evaluation metrics under the 80B-A3B MoE setup and with FP8 KV-cache storage when cache sizes are reported. At 1 million tokens, the paper reports prefill-FLOP reductions of 2.92 times against HySparse and 5.02 times against Hybrid SWA. Its listed FP8 cache sizes are 2.69 GB, 6.72 GB, and 12.09 GB, respectively, for HySparse2, HySparse, and Hybrid SWA. After light post-training of approximately 100 billion tokens, the reported mean MRCR-v2 gain was 11.30 points against HySparse and 6.44 points against Hybrid SWA; the reported mean RULER-v2 gains were 19.81 and 18.65 points. In the 256K-context RULER-v2 comparison, HySparse2 recorded 58.45. HySparse is listed at 32.61, while Hybrid SWA is listed at 35.74.

The paper also reports higher pretraining RULER and NoLiMa numbers for HySparse2 in its pretraining table, with RULER = 90.77 for HySparse2 versus 84.89 for HySparse and 88.71 for Hybrid SWA, and NoLiMa = 49.76 for HySparse2 versus 40.27 for HySparse and 30.13 for Hybrid SWA. The authors report that HySparse2 yields lower AgentPPL and LongPPL than the baselines across the evaluated context range up to 256K, but the accessible text does not present single headline percentage trajectories for those curves, and the paper does not equate benchmark curves with production costs or customer-facing latency.

TechNode included a rounded statement that HySparse2 cuts long-context costs and that cache footprints fall roughly 4.5 times at 1M tokens; that rounded summary differs from the paper’s tabulated FP8 cache figures and the paper’s explicit FLOP ratios. Other secondary outlets repeat similar high-level efficiency claims, but the arXiv preprint should be treated as the controlling document for exact numbers, baseline definitions, and experimental conditions.

HySparse2’s reported strengths are framed in the paper as improved long-context retrieval and multi-turn agentic task performance alongside reductions in prefill computation and KV-cache storage in the paper’s experimental setup. Those are paper-reported comparative outcomes on research models and must be qualified as such when connecting them to product plans.

The HySparse2 paper itself and TechNode’s reporting do not document a publicly available MiMo-V3 model or confirm HySparse2 deployment in any production MiMo-V3 instance. The Xiaomi MiMo release log opened during research does not list MiMo-V3, and the lack of an entry in that log is therefore a verified absence in the opened record rather than disproof of any unannounced or internal work.

HySparse2 is available on arXiv for detailed inspection and method reproduction; the arXiv metadata dates the v1 preprint to September 22, 2026 and identifies Xiaomi LLM-Core as an author affiliation. For readers who want the technical paper, consult the arXiv HTML version.

The paper-reported gains and Xiaomi’s previewed interest in applying HySparse2 to a future MiMo-V3 outline a research-to-product pathway, but the research metrics are tied to a specific 80B-A3B MoE benchmark configuration and to the experimental conditions the authors report. Any claim about MiMo-V3 shipping HySparse2 at scale or producing identical customer-facing improvements would require confirmation from a Xiaomi announcement or a documented release.

The reported MiMo-V3 preview indicates a possible Xiaomi research-to-product direction for long-context agent workloads, while the HySparse2 preprint supplies the benchmark measurements reported by TechNode; any product implementation remains unconfirmed. The relationship between the published research and an eventual product build will depend on further engineering validation, product decisions, and public disclosures from Xiaomi.