✨ 309B - 15B active ✨ MIT license ✨ Hybrid attention: 256k context with ~6× lower KV cache ✨ Generates responses ~3× faster with native multi-token prediction