| ▲ | 2001zhaozhao 37 minutes ago | |
I think this model will be a crazy big deal, it might be the first model that can reliably run multi-agent reasoning harnesses on local hardware, as this is a MoE model and can have several agents each with their own KV caches that fit into VRAM. Dethroning Qwen 3.8 is a big deal too, and I like their section about CoT monitorability. (This model beats Qwen on benchmarks - TML is not known to benchmaxx nearly as hard as Qwen.) I guess TML really are a bunch of rationalists who know what they're doing, sort of like Anthropic. | ||