851 2 weeks ago

Qwen 3.6, MTP-enabled, 512k context (20GB KV cache footprint with OLLAMA_KV_CACHE_TYPE=q8_0)