Fuli Luo

Now building @XiaomiMiMo. Previously @deepseek_ai

[Fuli Luo @_LuoFuli]

MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.

Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once.

Compared with MiMo-V2.6's Hybrid SWA architecture:
• 5.02× lower prefill FLOPs at 1M tokens
• 4.5× smaller KV cache at 1M tokens
• Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL

Why build a new architecture?
Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-cache size, and retrieval accuracy on the critical path at the same time.

HySparse2 tackles all three with two levels of KV sharing:
• KV Bridging: Following YOCO, full-attention layers in the cross-decoder build their K/V from self-decoder hidden states.
• KV Reuse: Within each hybrid block, sparse layers reuse the preceding full-attention layer's KV cache and selection indices.

Two more changes: token-level selection replaces block-level selection, and a forced window of recent tokens replaces the separate SWA branch, so local and global tokens share one KV cache. Since all cross-decoder KV caches now come from the self-decoder, prefill can stop once the self-decoder finishes.

Paper: https://t.co/REeEdd5pL7

[Fuli Luo @_LuoFuli]

@ValsAI Nice work.
Internally, our RL setup is hackable environments plus a separate anti-hack preprocessing script. The upside of that split is that when a new hack pattern shows up, we only update the script and do not rebuild the environments. The downside is that what we ship is not out-of-the-box: it has to be paired with the preprocessing script.

After you flagged this, our plan was to tighten the on-the-fly anti-hack scripts in mimoagent so we would not have to rebuild images. That said, @adithya_s_k's approach of cleaning the leaks and then rebuilding the images is more convenient for people running other harnesses.

https://t.co/xBSKVcHH9L

[Adithya S K @adithya_s_k]

We saw many of the same reward hacking issues when we tested these environments a week ago with our Harbor implementation of Mimo RL Envs, and patched most of them 3 days ago.

Please try breaking them and let us know what you find. We want to be fully transparent and share what we learn openly, including common reward hacking patterns and how to avoid them.

[Fuli Luo @_LuoFuli]

MiMo-V2.6: The Hard Road to Scaling Up RL

MiMo-V2.6 is very likely one of the largest single RL runs, by compute, that any open-source model team has undertaken to date. In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL. That takes more than research conviction. It takes a vision for AGI, respect for the unknown, and the nerve to walk straight into the hardest problems.

The result is a model whose potential was built through mid-training and unlocked through heavy RL. Today, it is the number one open-source model. I strongly recommend reading the technical report. I believe it will become one of those papers that Agent RL practitioners keep reopening and discovering something new in each time. In my view, the research innovations and engineering challenges behind it surpass those of DeepSeek R1, which I was partly involved in.

Some will ask: why MixRL instead of MOPD? First, they are not competing choices. We ran MixRL on verifiable tasks of moderate difficulty, including code and related agentic tasks, and found that the resulting models generalize remarkably well. Second, tasks that are difficult to verify, extremely long-horizon, or simply too challenging to include in a joint RL run are trained separately. Including them would substantially reduce rollout efficiency or introduce significant rollout staleness. We then merge the resulting capabilities through MOPD. Games, 3D tasks, and tasks with subjective evaluation signals all fall into this category.

There is also a third, slightly cheeky answer. Our team is flat enough and free enough of organizational silos that MixRL simply is not difficult for us. More importantly, everyone enjoys working this way. People from different domains come together every day, driven by the pursuit of AGI and intelligence that can continuously improve itself, to confront and resolve the RL bottlenecks in each field. I will always remember the RL daily update meetings from this period. They were intense and dense, with intelligence emerging in real time.

To help the open-source community focus on solving real Agentic RL problems, we have released a Qwen model distilled from MiMo RL trajectories as a stronger starting point for RL, along with 7K diverse environments and a complete RL training framework. We hope these resources will help move Agentic RL research forward.

MiMo-V2.6 is only the beginning. In an era when intelligence is easy to replicate, we still choose the hard road toward self-improvement and AGI. Much of what lies ahead remains unknown. But we are willing to keep investing the time, compute, and passion required to take on one hard problem after another and work each of them all the way through, until intelligence crosses into a new regime.

[Fuli Luo @_LuoFuli]

We hope this livestream sparks the research community’s interest in the core challenges of scaling RL and encourages researchers to help us refine our training recipe. We’ve been delighted to see so much insightful analysis based on the detailed training metrics.

[Fuli Luo @_LuoFuli]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.

Streaming the run: https://t.co/ZSxahzJRju

[Fuli Luo @_LuoFuli]

We believe RL is one of the most scalable and efficient paths toward self-improvement.

[Lei Li @_TobiasLee]

Our MOPD from MiMo-V2-Flash has been widely adopted in modern post-training pipelines.

Now the paper is out with more details & comparison.

Check it out: https://t.co/8XLugDt1cN

[Fuli Luo @_LuoFuli]

A strong model evolution needs a solid harness system, and vice versa. 14 days, 5 people, one vibe-coding journey — and MiMo Code was born. It's open source: https://t.co/Yb0aPX5IOH

[Xiaomi MiMo @XiaomiMiMo]

🚀 1,000+ TOKENS/S ON A 1T MODEL! 🚀

We are thrilled to release Xiaomi MiMo-V2.5-Pro-UltraSpeed in collaboration with @TileRT_AI , breaking the 1,000 tokens/s output speed on a 1 Trillion parameter model for the FIRST TIME!

Not wafer-scale integration like Cerebras. Not pure on-chip SRAM chips like Groq. We achieve 1,000 tps on a 1T MoE model using just a SINGLE, STANDARD 8-GPGPU NODE.

Read the full technical deep dive：https://t.co/MX0kjHKdKi

Want to experience the future of real-time AI?
👉 Apply for UltraSpeed now: https://t.co/aeWAxyhwVk
⏳ Limited-Time Access: Application-based · Jun 8 – Jun 23 (PDT)
💬 Chat Experience: Completely FREE for a limited time — try the blazing-fast web chat now.
⚡ UltraSpeed API: Just 3x the price for a ~10x boost in output experience.
🤝 Enterprise & Large-Scale Needs: [email protected]

[Fuli Luo @_LuoFuli]

Inference Optimizations Behind the MiMo-V2.5 Series API Price Reductions

Read the full technical blog: https://t.co/B5tp4tdnim

The V2.5 model family, including MiMo-V2.5 and MiMo-V2.5-Pro, is built on a Hybrid Sliding Window Attention (Hybrid SWA) architecture, which compresses KVCache storage to roughly 1/7 that of Full Attention. However, architectural advantages rarely translate directly into measurable gains in production serving. To realize these gains, we redesigned KVCache management, tiered caching, and the prefix-cache tree; addressed key challenges in SWA KVCache handling; and optimized scheduling as well as the Prefill/Decode pipeline.

Validated on real production traffic, these optimizations have increased effective KVCache capacity by nearly 5x, with server-side cache hit rates averaging 93%–95% across mainstream harness frameworks. Together with MoE configuration tuning and multimodal inference optimizations, they enable more efficient long-context inference and form part of what makes the recent API price cuts possible.

[Fuli Luo @_LuoFuli]

@leonliuzx yes

[Fuli Luo @_LuoFuli]

Behind the MiMo API Price Reduction:
The deepest price cut, up to 99%, is for Input (Cache Hit). The core reason is our inference framework now supports hierarchical KV cache optimization for SWA. Production inference engine tests show this optimization increases cached token capacity by 5x, equivalent to an 80% reduction in caching costs. Combined with Cache Read Overlap among multiple Full Attention modules in the Hybrid model, actual costs are further reduced.

Prices for Input (Cache Miss) and Output are also reduced by 60%-80%. This mainly benefits from the extreme 1:7 Full:SWA sparsity ratio brought by the model architecture (the prefill compute of the 70-layer MiMo-V2.5-Pro roughly equals a 10-layer GQA model). This kept our original inference costs well below the industry average, naturally leaving a 2x-3x profit margin in pricing. This price adjustment simply reflects our decision to pass these structural cost efficiencies directly to developers.

Operating at these newly reduced API prices, our production inference engine is running at near full capacity, and we can still essentially break even. We previously advised LLM companies not to "blindly cut prices" precisely because very few model architectures and inference optimizations can keep API costs from running at a loss. If more architectures that save compute and KV cache emerge, along with better inference Infra to drive down API costs, this will form an excellent virtuous cycle in the industry.

More crucially, affordable, high-performance model APIs will drive real, sustained, and at-scale inference demand. This upstream demand pulls forward the development of the entire AI infrastructure chain—including chips, servers, optical transceivers, PCBs, liquid cooling, power, energy storage, and data centers—serving as a strategic fulcrum for a systemic revaluation of AI hardware. In the long run, this injects more affordable and accessible compute into both training and inference pipelines, accelerating the parallel evolution of global AGI across multiple regions and technical routes.

For more technical details, we will release a detailed Blog post later.

[Xiaomi MiMo @XiaomiMiMo]

🚀 Better inference efficiency, lower costs, broader access.

MiMo-V2.5 Series API pricing is now permanently reduced — by up to 99% compared to previous pricing.
✨ Unified pricing across all context lengths.
MiMo Token Plans have also been upgraded:
• 5–8× more usable tokens at the same price
• Simpler and more transparent billing rules
🎁 As a thank-you to current users, all current Token Plan credits will be fully reset.
🎧 MiMo-V2.5-TTS remains free for a limited time.
⏰ Effective May 26 at 6:00 PM PDT.

These improvements are powered by continued inference optimization and serving efficiency upgrades across the MiMo stack.
🛠️ We’ll also publish a detailed technical blog on the inference optimizations later — stay tuned.

[Fuli Luo @_LuoFuli]

Just dropped two open-source models: MiMo-V2.5-Pro (Code Agent, 1T total) and MiMo-V2.5 (Multimodal Agent, 310B total).
Oh and one more thing — we're giving devs & creators 100T tokens on us. Go build something cool 🛠️
🎁 100T Free Token Grant for Builders https://t.co/Sfykf7As2J

[Xiaomi MiMo @XiaomiMiMo]

Xiaomi MiMo-V2.5 is now officially open-sourced！
MIT License, supporting commercial deployment, continued training, and fine-tuning - no additional authorization required.
Two models, both supporting a 1M-token context window :
• MiMo-V2.5-Pro: built for complex agent and coding tasks, ranking No.1 among open-source models on GDPVal-AA and ClawEval
• MiMo-V2.5: a native omni-modal model with strong agent capabilities
A model's value isn't measured by rankings alone — it's measured by the problems it solves.
Let's build with MiMo now!
🤗 Weights: https://t.co/w7wNtXj9Rm
📄 Blog: https://t.co/bPaaWMRI0g

[Lei Li @_TobiasLee]

Big week for model releases, and Claw-Eval is updating too.

MiMo V2.5 Pro now ranks 3rd, and MiMo V2.5 ranks 5th.
Next up: DeepSeek V4?
👉🏻 https://t.co/wVioDjWntX

[Xiaomi MiMo @XiaomiMiMo]

Xiaomi MiMo-V2.5 Series: Pushing Open-Source Agents Forward

🔸 MiMo-V2.5-Pro, our strongest model yet.
A major leap from MiMo-V2-Pro in general agentic capabilities, complex software engineering, and long-horizon tasks, now matching frontier models like Claude Opus 4.6 and GPT-5.4 across most benchmarks (SWE-bench Pro 57.2, Claw-Eval 63.8, τ3-Bench 72.9). It can autonomously complete professional tasks involving 1,000+ tool calls, work that would take human experts days.

Tech Blog: https://t.co/I275kxl9kL

🔸 MiMo-V2.5, native omnimodal with strong agentic capabilities.
Pro-level agent performance at roughly half the cost. Improved multimodal perception across image and video understanding, native 1M-token context window, and significantly more efficient inference.

Tech Blog: https://t.co/7DLo0n2MNu

🔗 API & Token Plan: https://t.co/P74SR2MaJe

[Fuli Luo @_LuoFuli]

A bigger problem: many third-party harnesses compress tool responses every 3 steps when approaching the context limit, leading to very low cache hit rates.

[Fuli Luo @_LuoFuli]

Two days ago, Anthropic cut off third-party harnesses from using Claude subscriptions — not surprising. Three days ago, MiMo launched its Token Plan — a design I spent real time on, and what I believe is a serious attempt at getting compute allocation and agent harness development right. Putting these two things together, some thoughts:

1. Claude Code's subscription is a beautifully designed system for balanced compute allocation. My guess — it doesn't make money, possibly bleeds it, unless their API margins are 10-20x, which I doubt. I can't rigorously calculate the losses from third-party harnesses plugging in, but I've looked at OpenClaw's context management up close — it's bad. Within a single user query, it fires off rounds of low-value tool calls as separate API requests, each carrying a long context window (often >100K tokens) — wasteful even with cache hits, and in extreme cases driving up cache miss rates for other queries. The actual request count per query ends up several times higher than Claude Code's own framework. Translated to API pricing, the real cost is probably tens of times the subscription price. That's not a gap — that's a crater.

2. Third-party harnesses like OpenClaw/OpenCode can still call Claude via API — they just can't ride on subscriptions anymore. Short term, these agent users will feel the pain, costs jumping easily tens of times. But that pressure is exactly what pushes these harnesses to improve context management, maximize prompt cache hit rates to reuse processed context, cut wasteful token burn. Pain eventually converts to engineering discipline.

3. I'd urge LLM companies not to blindly race to the bottom on pricing before figuring out how to price a coding plan without hemorrhaging money. Selling tokens dirt cheap while leaving the door wide open to third-party harnesses looks nice to users, but it's a trap — the same trap Anthropic just walked out of. The deeper problem: if users burn their attention on low-quality agent harnesses, highly unstable and slow inference services, and models downgraded to cut costs, only to find they still can't get anything done — that's not a healthy cycle for user experience or retention.

4. On MiMo Token Plan — it supports third-party harnesses, billed by token quota, same logic as Claude's newly launched extra usage packages. Because what we're going for is long-term stable delivery of high-quality models and services — not getting you to impulse-pay and then abandon ship.

The bigger picture: global compute capacity can't keep up with the token demand agents are creating. The real way forward isn't cheaper tokens — it's co-evolution. "More token-efficient agent harnesses" × "more powerful and efficient models." Anthropic's move, whether they intended it or not, is pushing the entire ecosystem — open source and closed source alike — in that direction. That's probably a good thing. The Agent era doesn't belong to whoever burns the most compute. It belongs to whoever uses it wisely.

[Xiaomi MiMo @XiaomiMiMo]

Introducing MiMo-V2-Pro & Omni & TTS
https://t.co/qdufHaBJWY

[Fuli Luo @_LuoFuli]

MiMo-V2-Pro & Omni & TTS is out. Our first full-stack model family built truly for the Agent era.

I call this a quiet ambush — not because we planned it, but because the shift from Chat to Agent paradigm happened so fast, even we barely believed it. Somewhere in between was a process that was thrilling, painful, and fascinating all at once.

The 1T base model started training months ago. The original goal was long-context reasoning efficiency. Hybrid Attention carries real innovation, without overreaching — and it turns out to be exactly the right foundation for the Agent era. 1M context window. MTP inference for ultra-low latency and cost. These architectural decisions weren't trendy. They were a structural advantage we built before we needed it.

What changed everything was experiencing a complex agentic scaffold — what I'd call orchestrated Context — for the first time. I was shocked on day one. I tried to convince the team to use it. That didn't work. So I gave a hard mandate: anyone on MiMo Team with fewer than 100 conversations tomorrow can quit. It worked. Once the team's imagination was ignited by what agentic systems could do, that imagination converted directly into research velocity.

People ask why we move so fast. I saw it firsthand building DeepSeek R1. My honest summary:

— Backbone and Infra research has long cycles. You need strategic conviction a year before it pays off.
— Posttrain agility is a different muscle: product intuition driving evaluation, iteration cycles compressed, paradigm shifts caught early.
— And the constant: curiosity, sharp technical instinct, decisive execution, full commitment — and something that's easy to underestimate: a genuine love for the world you're building for.

We will open-source — when the models are stable enough to deserve it.

From Beijing, very late, not quite awake.