SK Telecom opens A.X K2: 688B, Apache 2.0, 97% on AIME26
Korea's state model contest just shipped two 700B open giants in a week; on Aug 8 a government cull thins four teams to three.
Copy markdown
33B active under a real Apache 2.0 license
A.X K2 is a 688B-parameter Mixture-of-Experts that fires just 33B params per token, and it ships under Apache 2.0 — so you can put it in production, not just a research demo. Thinking-mode scores hit 97.1% on AIME26, 85.6% on GPQA Diamond, and 98% on the τ²-Bench telecom-agent test, with a 256K context (native 128K, YaRN-extended). Weights are live on Hugging Face and GitHub now.
Strong at reasoning, thin as a web agent
SKT claims a 32.2-point average jump over its 519B A.X K1 across 14 benchmarks and calls K2 comparable to Qwen3.5 and DeepSeek-V4 on reasoning. But agentic breadth is uneven — it aces telecom tool-use yet reportedly lands near single digits on open-web BrowseComp, and 688B total still needs serious GPU memory even when sparse. Reach for it on reasoning and Korean-STEM work, not as a drop-in autonomous browser.
Motif-3 cracks the open top 3 — but research-only
A ~30-person startup, Motif, put a 314B/13B-active MoE at roughly 3rd among open-weight models on the Artificial Analysis index (44), beating models many times its active size. The catch for builders: the license is non-commercial research only — a step back from Motif-2's Apache terms — with a production-ready final targeted for early August. Watch it, don't ship it yet.
Why the flood: a state elimination round on Aug 8
These aren't coincidences — they're contest entries. Korea's Ministry of Science runs a sovereign-model program (LG, SKT, Upstage, Motif), and an Aug 8-11 evaluation reportedly cuts the four teams to three; the survivors power a national chatbot offered free to ~51M citizens and must run at least half of it on certified domestic models. Expect more open weights to drop before the deadline.
Elsewhere: the rest of the Korean open-weights wave
The other 700B in this week's pair, LG's EXAONE 2.0, is also Apache 2.0 and tuned for long-context retrieval. Upstage's agent-focused Solar Open 2 (250B, ~15B active) claims frontier scores while running on just two GPUs. Between them, that's a rare cluster of commercially-usable frontier open weights you can pull this weekend.