The company behind PUBG just trained its own 29B Local AI model from scratch. And the architecture is interesting. KRAFTON’s new Raon-1 mini is … 🧠 28.8B total parameters ⚡ only 4.3B active/token 💾 official Q4_K_M → 19.7GB 📚 262K context 🔓 Apache 2.0 🤖 reasoning + agent/tool use 18 of its 25 layers use fixed-state linear attention. Only 7 layers maintain a growing KV cache. KRAFTON says those layers use only ~8 KiB of BF16 KV per token. So even at the full 262K context, that’s roughly: 👉 ~2GB of growing KV cache before other runtime/state overhead. And in KRAFTON’s reported comparisons, this new 30B-A4B model beats Qwen3.6-35B-A3B on 11 of 18 listed benchmarks. Lots of new models coming for Qwen3.6-35B recently. ⚠️ Preview release. KRAFTON’s benchmark results. Upstream llama.cpp support is still pending and KRAFTON provides its own branch today. Link in ALT