Xiaomi Officially Announces MiMo-V2.5 Will Soon Be Open Source, Going Head-to-Head with DeepSeek

Last night, Xiaomi suddenly announced that the MiMo-V2.5 series models are entering public beta—and more importantly, MiMo-V2.5-Pro and MiMo-V2.5 will soon be open-sourced globally. This marks another major move by Luo Fuli, Xiaomi’s MiMo large model leader (formerly a core member of DeepSeek), just 36 days after the V2 series launch.

The timing is intriguing. DeepSeek V4 is rumored to be released this week, yet Xiaomi moved first, releasing two open-source models with parameter scales and positioning that directly benchmark the DeepSeek series. Developers are already discussing: is this a head-to-head battle on the open-source track?

Two Models, Each with a Focus

The core of this release lies in two models:

MiMo-V2.5-Pro is Xiaomi’s strongest current model, specializing in long-chain agent tasks. Official data shows it can stably perform nearly a thousand rounds of tool calls, achieving 73.7 points on the MiMo Coding Bench—narrowing the gap with Claude Opus 4.6 (77.1 points).

Real-world cases illustrate its power. When tasked with implementing a complete SysY compiler in Rust (a Peking University “Compiler Design” project that normally takes undergraduates weeks), V2.5-Pro finished in 4.3 hours using 672 tool calls, scoring a perfect 233 points on the hidden test set. The process included lexical and syntax analysis, AST, Koopa IR generation, RISC-V backend, and performance optimization. With a cold start success rate of 59%, it means the architecture was right on the first try.

Another case involved developing a web app video editor. V2.5-Pro wrote 8,192 lines of code, completed 1,868 tool calls in 11.5 hours, and delivered an app featuring multi-track timelines, clip trimming, crossfades, and audio mixing.

Even more technical: it designed an FVF-LDO (Flip Voltage Follower Low-Dropout Linear Regulator) based on TSMC’s 180nm CMOS process. Experienced analog circuit engineers need days for this, but V2.5-Pro, connected to an ngspice simulation loop, generated a compliant design within one hour—improving four key metrics by an order of magnitude.

MiMo-V2.5 is a general-purpose, multimodal agent model capable of handling images, audio, and video simultaneously, with faster inference speed than the Pro version, making it suitable for latency-sensitive tasks. On Claw-Eval (an end-to-end benchmark for AI agents), V2.5 surpasses the previous flagship V2-Pro and reduces API cost by about 50%.

In multimodal evaluations, V2.5 approaches or even surpasses Claude Opus 4.6, Gemini 3 Pro, and GPT-5.4 on VideoMME, CharXiv, and MMMU-Pro. For coding, it exceeds Gemini 3.1 Pro on everyday tasks, though it still trails Claude Opus 4.6.

Token Efficiency: The Secret Weapon

Performance aside, token efficiency is the more practical advantage.

Xiaomi’s comparison data shows:

• MiMo-V2.5-Pro saves 42% tokens over Kimi K2.6 for the same Claw-Eval score.
• MiMo-V2.5 saves 50% tokens compared to Meta’s Muse Spark.

What does that mean? For the same task, running it with MiMo cuts the cost in half. For agent applications requiring frequent API calls, that difference can multiply dozens or hundreds of times.

Xiaomi also optimized its Token Plan subscriptions:

• Removed the old 1 Token = 4 Credits pricing.
• No longer distinguishes between 256k and 1M context windows.
• Added a nightly discount period (00:00–08:00), reducing Credit consumption by 20%.
• Introduced monthly and annual subscriptions, with up to ¥948.96 off for annual plans.

Billing now is 1x (1 Token = 1 Credit) for V2.5, and 2x (1 Token = 2 Credits) for V2.5-Pro. After the Token Plan launched, many users complained prices were high and cheaper plans insufficient—this adjustment seems to be Xiaomi’s response to market feedback.

The Logic Behind the Open-Source Strategy

Why is Xiaomi going open source?

Luo Fuli said during the previous model release that “the models will be open-sourced once they become stable enough.” That promise is now fulfilled—at a tactically subtle time, right as DeepSeek V4 looms. Xiaomi released two open-source models first, positioning both to mirror DeepSeek’s lineup.

In March, Xiaomi’s MiMo-V2-Pro appeared anonymously on OpenRouter as “Hunter Alpha,” and developers even mistook it for DeepSeek V4. Now, the V2.5 series launches in the same week as rumors of DeepSeek V4—perfectly timed.

What are the advantages of going open source? The most direct is capturing the developer ecosystem. In the agent era, whichever model developers use most frequently will dominate the infrastructure narrative. As a hardware company, Xiaomi’s self-developed large models have system-level permissions and deep integration with its services—open-source models can accelerate full-scenario implementation across its “human–vehicle–home” ecosystem.

Another consideration is cost. Open-source models allow local deployment, meaning scenarios requiring massive calls (like Xiaomi’s own IoT devices or cars) don’t need repeated cloud API requests—saving huge inference costs.

Market Reaction: Optimism and Skepticism

The developer community is divided on MiMo-V2.5.

Optimists focus on token efficiency and Xiaomi’s open-source promise. Some users, who have long used MiMo in tools like Hermes (Cursor’s nickname), remark that “it’s free anyway, performs about the same as other Chinese models, and its writing ability is unbeatable.” With open source, opportunities for local deployment and retraining open up—real value for teams needing customization.

Skeptics question market acceptance. One bluntly said, “No one cares. I’ve never heard of anyone using Xiaomi’s models.” While harsh, it reflects a reality—Xiaomi entered the large-model race late, with weaker brand recognition than DeepSeek, Kimi, or Qwen. How much Luo Fuli’s “DeepSeek halo” benefits Xiaomi depends on results to come.

Others note Luo’s earlier remark that “cheap tokens don’t make money.” Now Xiaomi is dropping prices and going open source—so is this a strategic loss-leader to grab market share, or has it truly optimized costs? The answer will likely come after open source goes live and community results emerge.

Technical Details: Agent Capabilities at the Core

MiMo-V2.5 series centers around agent capabilities.

V2.5-Pro, capable of stable thousand-round tool calling, ranks at a top-tier level. The hardest part of long-chain reasoning is maintaining logical consistency—ensuring context from hundreds of prior turns is accurately referenced in later ones, while managing tool failures and intermediate anomalies. Xiaomi’s internal tests show significantly improved instruction adherence, with the model capturing implicit intents and staying on track over extended interactions.

Multimodal perception is another highlight of V2.5. It processes images, audio, and video simultaneously, and acts on what it perceives. This capability shines in tasks like having an agent watch a tutorial video and then operate software accordingly, or listen to a voice instruction while interpreting a screenshot to grasp user intent.

In programming, V2.5 surpasses Gemini 3.1 Pro in daily coding tasks but remains behind Claude Opus 4.6 in complex refactoring and large-scale architectural design. Still, for most developers—writing code, fixing bugs, generating test cases—V2.5 is more than capable.

After Open Source: What to Watch For

Xiaomi says MiMo-V2.5-Pro and MiMo-V2.5 will be “open-sourced globally soon,” but hasn’t given a timeline. Typically, releasing weight files, code, and training details takes weeks to months.

After open source, developers will focus on several key aspects:

1. Model size and hardware requirements. If it’s too large, regular users can’t run it, making open source lose meaning. After DeepSeek V3’s release, many complained about high VRAM demands; will Xiaomi also release quantized versions?
2. Training data and methods. Luo Fuli was known for transparency at DeepSeek—will MiMo follow suit and disclose training details? If only weights are released but not data/methods, community exploration is limited.
3. Commercial licensing. Open source doesn’t always mean free for commercial use. The licensing (MIT, Apache 2.0, or custom) will directly affect enterprise adoption.
4. Ecosystem tools. Releasing the model is just the start—will inference frameworks, deployment tools, and fine-tuning scripts follow? DeepSeek’s open source quickly spawned optimized variants and implementations; can Xiaomi emulate that success?

Final Thoughts

Xiaomi’s move was swift, but whether it can stand firm in the open-source model race depends on execution.

DeepSeek’s success isn’t just technical—it’s also about transparency and community engagement. Luo Fuli brings DeepSeek’s DNA to Xiaomi; whether she can replicate that approach will be pivotal.

Another variable is Xiaomi’s hardware ecosystem. If MiMo deeply integrates with Xiaomi’s phones, vehicles, and IoT devices, forming a coordinated cloud–edge agent infrastructure, then open-source models become more than technological outputs—they become ecosystem foundations. OpenAI and Google are both pursuing this path, but Xiaomi, as a hardware manufacturer, holds a unique advantage.

As for the head-on DeepSeek showdown—it’s too early to tell. V4 hasn’t launched yet, and V2.5 isn’t open-sourced. Both sides are holding their best cards. Developers hope for healthy competition driving progress—stronger models, lower costs, and deeper open source.

From that perspective, Xiaomi’s entry is a good thing.

---

References

• Xiaomi MiMo Going Open Source! – Linux.do Community Discussion – Developers’ first reactions and discussions on MiMo’s open source plan
• Before DeepSeek V4 Launch, Luo Fuli Makes Her Move: Xiaomi’s MiMo-V2.5 Midnight Surprise – 36Kr – Technical interpretation and case studies
• Xiaomi Releases MiMo-V2.5 Series Models – YouTube AI Morning Brief – Video report
• Xiaomi Large Model MiMo-V2.5-Pro and MiMo-V2.5 to Go Global as Open Source – PChome – Official release summary