Solar Open 2, an open model that reads a million words on about a quarter of the memory
Solar Open 2
A long AI conversation gets expensive for one big reason. The model has to keep hold of every word you have said. Upstage, an AI lab in South Korea, just released Solar Open 2. It is a very large open model that reads about a million words, a whole shelf of books, on roughly a quarter of the usual memory.
250 billion internal numbers, only 15 billion awake per word
Solar Open 2 is a big set of specialist sub-networks, and for each word it wakes only a few of them. So it holds two hundred and fifty billion internal numbers in total, but only about fifteen billion do the work on any one word. So a model this large runs at a small model's cost. The weights are public and free to download. And it is built for real work with tools, code, and documents.
Only 1 layer in 4 stores every past word
That expert setup is not the new part. What is new is how Solar Open 2 handles memory. A model runs each word through dozens of steps, called layers. In a normal layer, to write the next word, the model looks back over every word it has stored so far. Rereading that whole pile each time is what makes long chats slow and expensive. Solar Open 2 lets only one layer in four keep that pile. So what are the other three doing instead?
Three of four layers keep a running summary, not the words
They keep a small running summary instead. As the model reads, each of these layers holds one fixed-size note and updates it with every new word. The real name for this is linear attention. Because that note never grows, these layers store nothing extra per word. That is what lets a million-word context stay affordable. And this note can do two things. It can write new facts in, and wipe old ones out. Picture a whiteboard you can write on and erase, not a notepad where you only keep adding pages. That erasing is what keeps one small note useful even after a million words, instead of letting it clog with stale facts.
Only 12 of 48 layers store the past
Solar Open 2 also drops one thing almost every other model keeps, its position tags. Most models stamp each word with a number for its place in line, so the model can tell the order. Those tags are also what break when you push a model past the length it was trained on. Solar keeps none. It can, because the running-summary layers read strictly left to right, so the note already carries what came before. Dropping the tags is what stretches the window to a million words. Only twelve of the model's forty-eight layers still keep that growing pile of words. So its memory over a long input is about a quarter of a normal model the same size.
Leads a couple of tests; two bigger rivals win more
Every score here is Upstage's own testing, against other open models its size or bigger. Solar Open 2 leads a few. It tops MMLU Pro, a broad exam of knowledge and reasoning, at eighty-six point two. And it edges a coding test called LiveCodeBench, ninety-two point four to ninety-two point three, close enough to a tie. On many other tests, two larger rivals finish ahead, DeepSeek V4 Flash and MiMo. So Solar Open 2 holds its own against the strongest open models while being smaller, but it is not the outright leader. These numbers are Upstage's own, the only run of them so far.
Free for commercial use, if you name derivatives Solar
You can download the weights right now, and the full technical report is on arXiv. The license lets you use Solar Open 2 commercially. But if you build your own model on top of it, you have to name that model Solar and display the words Built with Solar. So these are open weights, not open source.
A million-word memory, kept as a running summary
Keep a running summary in place of every word
Most of what makes a long AI conversation expensive is the model holding on to every single word. Solar Open 2 keeps a short running summary on three of every four layers instead. That is how a model this large reads a whole shelf of books at once, on about a quarter of the memory. And you can download it today.
