⇄ Korean LLM comparison
A.X 4.0 vs Claude
Independent attribute comparison. A paid diagnostic measures these on your own 20 prompts.
✦Independent · zero vendor funding·Data verified 2026-07-13·Methodology & sources ↗
| A.X 4.0 | Claude | |
|---|---|---|
| Vendor | SK Telecom | Anthropic |
| Params | 72B / 7B Light | 비공개 |
| Context | 128K | 200K |
| License | Open weights · commercial (Qwen License, MAU < 100M; 7B Light is Apache 2.0) | API-only |
| Pricing | Open weights, self-host / on-prem. Based on Qwen2.5 with added Korean training. | Varies by tier (premium tiers costly). Korean token efficiency may trail Korea-specialized models. |
| Korean strength | KMMLU 78.3 (> GPT-4o 72.5); ~33% better Korean token efficiency than GPT-4o. | Top-tier Korean long-form instruction-following among global models. |
| Best for | On-prem enterprise, Cost-sensitive Korean work, Data-sovereignty needs | High-quality complex tasks, Long-form analysis, Careful instruction-following |
✦ How to decide
- ›License & hosting: A.X 4.0 = Open weights · commercial (Qwen License, MAU < 100M; 7B Light is Apache 2.0); Claude = API-only.
- ›Cost depends on your Korean token volume and token efficiency, not list price alone.
- ›Korean quality (honorifics, refusals, long-form) varies by use case — measure, don't assume.
Decide on your own prompts in 48h
48h Diagnostic ₩490,000 · tax invoice available
More comparisons