Xiaomi releases MiMo-V2-Flash, an open-weight MoE model with 309B total and 15B active parameters, saying it excels in reasoning, coding, and agentic scenarios
improve math, break coding. Enhance reasoning, hurt safety. ✅ Solution: Train specialized expert [image] Elie / @eliebakouch : wow, this looks like a very solid open model by Xiaomi, competing with K2/DSV3.2 on benchmarks with fewer parameters. it's MIT licensed, with a very good tech report and base/thinking versions available it's using the same sliding window attention arch as gpt-oss (sink with SWA [image] Bill Gurley / @bgurley : Wow! @nvidia & @Xiaomi both dropped open models in past 24 hours. There is a non-zero chance open models dominate the vast majority of tokens in the long run. Lot's of players interested in that future. I would argue that Amazon, Apple, and Meta should be also. Not sure they do. @xiaomimimo : 🏗️ Architecture Deep Dive MiMo-V2-Flash uses hybrid attention: Sliding Window + Global in a 5:1 ratio → 8 hybrid blocks (5 SWA layers → 1 Global) → 128-token sliding window → 256 experts, 8 active per token → Natively 32K trained, extended to 256K context support 💡 Key [image] @openrouterai : Congratulations to the @XiaomiMiMo team on the release of MiMo-V2-Flash! MiMo-V2-Flash is live now on OpenRouter, free for a limited time. Zack Angelo / @zackangelo : interesting bit from the @XiaomiMiMo technical report: a small 128 token sliding window outperformed a larger 512 token window as well as global attention everywhere [image] @lmsysorg : 🚀 Day-0 support for @XiaomiMiMo Mimo-v2-flash in SGLang! We're excited to announce day-0 support for Mimo-v2-flash in SGLang. Mimo-v2-flash is an inference-centric model featuring sliding-window attention (SWA) and multi-layer MTP: Near-zero-overhead multi-layer MTP support @teortaxestex : OK, Lei Jun officially has an open AGI lab. Have people started caring yet? I recall some telling me that Xiaomi MiMo will be limited to «on-device VLMs» or some such BS. No. You still don't get it. Post-DeepSeek China is different. They were shamed into getting serious. [image] @tphuang : Xiaomi has worked on LLMs for a while, mostly to improve performance on its phones. Now as ByteDance has rolled out Doubao phone, it is more important than ever for Xiaomi to also have a competitive LLM to carry out Agentic task in same way on phones, pads & EVs. China is a very @xiaomimimo : 🔧 Day-0 Open Source Inference Support All inference code is now available in SGLang — fully open source from launch. Community benchmarks by @lmsysorg on H200 (single node): 📊 Prefill: ~50K tokens/sec | TTFT < 1s 📊 Decode (3-layer MTP, 16K ctx): 5K-15K tokens/sec | 115-151 [image]
Context & Ripple Effects
MiMo-V2-Flash extends Xiaomi’s earlier open-source MiMo reasoning-model push from a model launch into a larger open-weight lineup aimed at reasoning and developer use cases. The MIT license and immediate support through SGLang and OpenRouter make the release more than a benchmark claim: it is deployable through existing inference channels.
Related coverage later tracks Xiaomi carrying this strategy into MIT-licensed MiMo-V2.5 releases and a dedicated open-source coding assistant, while Z.ai has also positioned open models on coding performance and cost. MiMo-V2-Flash is therefore a meaningful waypoint in competition around efficient, accessible Chinese open weights.
First-order effects
- Developers can self-host or access an MIT-licensed MoE model with 15B active parameters, while SGLang and OpenRouter users gain day-one routes to evaluate it for reasoning, coding, and agentic workloads.
- Xiaomi gains a concrete distribution and ecosystem foothold for MiMo-V2-Flash, though its performance and safety advantages remain claims that users will need to validate in their own workloads.
Second-order effects
- Rival open-model providers face greater pressure to pair model releases with permissive licensing, inference-stack support, and credible efficiency claims—not just benchmark results.
- Inference providers and developer platforms can add another large open-weight option, increasing the value of optimized serving for sparse MoE architectures and long-context workloads.
Third-order effects
- If repeated releases keep pairing competitive capability with open licensing and deployability, model differentiation may shift toward inference efficiency, tooling, and specialized agents rather than exclusive API access alone.
- The trade-offs noted around reasoning, coding, and safety suggest that open-model adoption will increasingly depend on workload-specific evaluation and governance, not aggregate benchmark leadership.
The trend: Open-weight AI competition is moving from standalone model announcements toward efficient, deployable model families supported by inference ecosystems and specialized developer tools.