Xiaomi open-sources MiMo-⁠V2.6-⁠Pro, after a $2.6 million reinforcement-learning run

Near the top of the Hacker News front page, with about eleven hundred points: Xiaomi open sourced MiMo V two point six Pro, its most capable model to date.

Xiaomi released two open weight models, a larger Pro and a cheaper Flash, both handling text, images and audio natively. Pricing did not move. Their claim: Pro is the strongest open source model on the Artificial Analysis index.

Pro's reinforcement learning run cost about two point six million dollars of compute: thirty steps over roughly seven hundred fifty thousand trajectories, in under six days. Flash's cost about eight hundred fifty thousand dollars. Xiaomi streamed both runs live, and open sourced the code and the training environments.

On DeepSWE, a software engineering benchmark held out of training, Pro's score rose from about fifty eight points to seventy three. Flash's rose from about forty nine points to sixty six.

Commenter rao-v picks out the openness rather than the scores. The dashboard ran while the training did, and the report even includes the benchmarks where they did poorly. Alice G G's essay is called Attention is all you have.

An essay that opens with the Tetris effect: play it daily for a few weeks and the shapes start appearing in clouds and buildings. Whatever you focus on shapes how you think. And a recommendation engine picks that focus for you.

Commenter tripleee offers a blunter version, and this is their view rather than anything the essay argues. Peak internet, they reckon, was the short window when it was there but you were not permanently attached to it. So what did Sun get wrong?

Bryan Cantrill distills it to one line: Sun had become bored with the mechanics of running a business. A startup on OpenSolaris wanted to buy Sun hardware and could not get Sun to answer the phone. Dell answered an overnight web form by morning.

Commenter coreyh14444 remembers it from the customer's side, which is the same complaint the post makes. Buying from the old workstation vendors meant live sales meetings and endless quote revisions.

An explainer out of Georgia Tech runs a small GPT-2, about a hundred twenty million parameters, inside your browser tab. Type a prompt and watch the tokens, attention weights and next token probabilities move. An old model, chosen because it has the same parts as the current ones.

Commenter mhl47 says what it unstuck, and the page bears it out. Attention and the feed forward layer alternate through twelve stacked blocks, so other tokens mix back in at every one.

Kev is a family of small, openly licensed models for the dull decisions: which team owns this ticket, does it need a human, how annoyed is the customer. On questions it never trained on, the largest lands within one point of Jev, the hosted model it imitates.

Commenter nico pushes a smaller option, from their own experience rather than anything the project measured. For plain classification, fifty to a hundred labelled examples and a classifier over embeddings trains on an ordinary processor in under five minutes.