Upstage's Solar-Open2-250B posts frontier scores at 15B active

A Korean lab quietly shipping a model that lands near the top of open-weight benchmarks is the kind of thing worth stopping on. It is the second sizable Korean MoE we've covered this week, after Motif's 314B Motif 3. Upstage's Solar-Open2-250B model card describes a 250 billion parameter Mixture-of-Experts that activates only 15 billion parameters per token, and the interesting choice is not the size but what surrounds it.

The attention stack interleaves three linear-attention layers for every one softmax layer, repeated twelve times across 48 total layers, with no rotary position encoding at all. That is the piece of the recipe that lets the model claim a 1 million token context window without the usual quadratic cost, and it is the piece other labs are likely to study closely. The MoE side runs 320 routed experts plus one shared expert, activating eight routed plus the shared one on each token. Training reportedly ran through roughly 12 trillion tokens on NVIDIA B200 GPUs at a stated 2 million GPU hours, across English, Korean and Japanese.

The scores Upstage lists are competitive with the strongest open weights on paper: 86.2 on MMLU-Pro, 86.3 on GPQA-Diamond, 92.4 on LiveCodeBench v6, 70.4 on SWE-Bench Verified, and 97.7 on Ko-AIME'25. Those are self-reported in the model card and accompanying technical report, so treat them as the vendor's claim rather than settled ground until independent evaluations land.

Two constraints sit alongside those numbers: the license and the hardware. This ships under a custom Upstage Solar License that requires derivatives to prefix the 'Solar' name and display 'Built with Solar' attribution, which is not the Apache-flavored openness enterprise legal teams tend to expect. And the minimum recommended deployment is four H200 GPUs, with eight suggested; these are open weights for well-resourced teams, not laptops. What the model card does not spell out is the serving cost at full 1M context, or how the reasoning_effort mode capped at 131,072 thinking tokens trades quality against latency.

If the numbers hold up under third-party runs, the bigger signal is that a hybrid-attention MoE recipe from outside the US majors is now close enough to the frontier that Korean and Japanese teams have a serious home-language option, and everyone else has a cheaper long-context template to copy.