Xiaomi publishes a live post-training RL dashboard for MiMo v2.6, showing benchmark scores step by step

This is pretty neat. What would be a good reason for the other model providers not to do this?

Speculating here, but I assume researchers can estimate the size of closed models from training time, training speed and token counts. Anthropic and OpenAI also want to keep each other on their toes so neither ends up on the wrong side of another release race.

I didn't know two thirds of the training data would be source code.

That tag list is the RL run, not the pretraining run.

This is crazy, but sadly Anthropic and OpenAI will never do this. How did we get to a world where Chinese companies are more open than US or even EU companies?

Why are they doing this? To head off distillation accusations?

Sometimes you're confident about what you're doing and show how you work to the world. Keeping the garage door open, or at least making it translucent, is always cool.

China's official policy now prefers open models and open model development, that may be part of it. The bottom of the page says "Open is what we value."

When you run benchmarks while training, isn't that contamination? I'm genuinely not sure if this is normal in big labs now.

They evaluate checkpoints during training, they don't train on the benchmarks. That's common practice for big reinforcement learning runs.

Kinda yes. The benchmarks become a validation set, so models overfit to them a bit if they drive stopping criteria. Much less than using them in the training data, but it does induce some benchmaxxing and partly explains why new models always eke out gains.

They also exist to detect degradation. Datasets are not perfect, and a batch with too much bad data can ruin a run, so it's a chance to find bad data and improve filtering.

2.6 Pro started 2026-09-15 10:32 UTC. For some reason I thought training took much, much longer than the progress bar suggests. I'm on mimo 2.5 pro now and it's great for the price, hopefully the next one is multimodal.

Those are post-training reinforcement learning steps, not the full pretraining run.

I've used MiMo-V2.5 as my main software engineering model across a bunch of projects and the ROI is great. I hit hallucination loops once or twice, but a stop and continue fixes it, and the cost is an order of magnitude better than what I was paying Anthropic for similar quality.

2.5 Pro or the regular 2.5? I've always found the MiMo models really good at tool calling and following instructions.

I found mimo v2.5 fine for basic things like a one-off python script, but dumb compared to qwen 3.8-flash-next, and definitely not in the same class as GLM 5.2 or 5.3. It's fast but makes basic mistakes that only get caught later.

You'd think they would make it less obvious that they run their whole operation with Claude.

If you mean the UI style, definitely not Claude. It couldn't write a clear sentence like "what each step's samples are made of", it would use all-caps everywhere with more padding and gradients.

That DeepSWE benchmark puts it in frontier territory. The transparency here is what we need.

For reference, Mimo-v2.5-Pro scored 19% on DeepSWE 1.1, and Fable is at 70%, Kimi K3 at 69%, Astra at 74% on max effort. 2.6-pro just reached 63.7% by step 10, and even flash reached 60.7% by step 12.

I've been testing their next model for a week and it's a clear step up. 2.5-pro worked like a forgetful senior engineer new to my project: capable but needing steering once projects got big. The next one is much better at multitasking and more ambitious about gaps and next steps.