Xiaomi is training MiMo v2.6 in the open. Here is why that matters.

Xiaomi just did something I have not seen before. They launched a live reinforcement learning dashboard for their MiMo v2.6 models. Not a blog post. Not a paper release months after the fact. A real-time training board that shows you every step, every cost, every metric, as it happens.

Two models are training right now, and you can watch.

MiMo v2.6 Pro is at step 15. They have spent $1.04 million so far. 30.2 billion tokens consumed. 351,000 samples trained. DeepSWE benchmark score: 65.78.

MiMo v2.6 Flash is at step 20. $470,000 spent. 46.2 billion tokens. 477,000 samples. DeepSWE score: 60.77.

Both are still training. Both are improving. Pro gained +0.051 on their primary metric since step 1. Flash gained +0.112. You can watch the numbers move if you refresh the page.

What is actually on the dashboard

This is not a marketing page dressed up to look technical. It is an operational training board. The kind of thing an ML engineer would have on a second monitor.

You can see live step progress with acceptance rates. Cost tracking per run. Token counts per step and cumulative. Benchmark scores updated as training goes. A batch composition breakdown: 67.7% code, 13.1% visual, 12.1% general, 4.1% cyber, 3.0% chat. Infrastructure metrics like gradient norms, entropy loss, policy loss, KL divergence. Training logs with a live feed of accepted and rejected samples.

They even log their failures. One entry reads: "we restarted the flash run from step 15. reason: a type of infra error on one of datasets was not correctly detected over the past ~3 hours."

That kind of honesty is rare.

Why this matters

Most AI labs publish results after training is done. Some publish papers months later. A few release model weights. Almost nobody shows you the training process live, costs and failures included.

Xiaomi is doing this while spending over $1.5 million on compute for these two runs alone. The Pro run trains on a batch size of 1,568 sequences across 16 parallel runs. The Flash run processes over 9,800 samples per step. This is not a toy experiment.

Three things stand out to me.

First, confidence. If you show your work live, you believe the work is good. Xiaomi is not hiding behind curated results. They are letting you watch the parts that do not work too.

Second, transparency. Open-source model weights are common now. Open-source training data is rare. Open-source training process in real time is almost unheard of. Xiaomi is going further than most labs.

Third, focus. 67.7% of their training data is code. This is not a general-purpose chatbot. This is a model being built to write, review, and debug software. The DeepSWE benchmark they track is a software engineering benchmark. Xiaomi is building a coding agent, and they are showing you exactly how.

The numbers

| Metric | Pro | Flash | |--------|-----|-------| | Cost so far | $1,037,787 | $470,211 | | Tokens total | 30.2B | 46.2B | | Samples trained | 351k | 477k | | DeepSWE score | 65.78 | 60.77 | | Improvement since step 1 | +0.051 | +0.112 | | Step time | 2h 54m | 2h 26m | | Context length | 106k tokens | 124k tokens |

Flash is cheaper, faster, and processing more tokens per step. Pro is scoring higher on benchmarks. Both are still running.

What this means

If you are building AI products, especially coding tools, pay attention to this. Xiaomi is not just releasing a model. They are releasing a process. The training board at mimo.xiaomi.com/rl is a live record of how a modern AI model is built, from raw compute to benchmark scores.

Whether MiMo v2.6 ends up competing with Claude, GPT, or Gemini is almost beside the point. The precedent matters more: a major tech company training a frontier model in the open, showing every step, every dollar, every failure.

That is the kind of transparency the AI industry talks about but rarely delivers. Xiaomi is actually doing it.

---

Live dashboard: mimo.xiaomi.com/rl