SK Telecom drops A.X K2: 688B open-weights MoE on Hugging Face
Korea's telco leaps 519B→688B with +32.2pt benchmark gains and long-context wins — its open-weights answer to DeepSeek, Kimi and GLM.
Copy markdown
What dropped: 688B params, open on Hugging Face
SK Telecom released A.X K2, a 688-billion-parameter MoE — up from 519B in A.X K1 — with the weights posted to Hugging Face. You can pull, self-host or fine-tune it today; SKT calls it its strongest Korean-language open model yet.
The jump: +32.2 points across 14 benchmarks
Averaged over 14 Korean and international benchmarks, K2 lands 32.2 percentage points above K1. Long-context and agent evaluations rose about 83.9 points — the gains sit exactly where agentic workloads live.
Sparse Gate Attention, tuned for long context
K2 introduces a new Sparse Gate Attention design aimed at long-document, multi-step reasoning, with beefed-up math, science and Korean knowledge. If your app chokes on context length, it's worth a benchmark.
Where it lands in the open-weights race
Reviewers slot K2 alongside DeepSeek V4, GLM and Kimi — now with a telco-backed Korean entrant in the trillion-ish-scale open tier. For builders that's another second-source model to hedge against any single lab's license or access changes.
Elsewhere: Cursor opens a ₹649/mo India tier
Cursor launched 'Cursor Start,' a ₹649/month plan for Indian developers with UPI billing, Grok 4.5 and Composer, cloud agents, the iOS app and MCP support — a much cheaper on-ramp to agentic coding in a massive dev market.