Models we advise on

Picked from the world agent leaderboard, filtered to open-source (MIT / Apache / open licenses), all around or above 100 billion parameters — the scale where a serious cluster is required.

MIT

GLM 5.2 Max

Z.AI

A frontier-scale open-weight reasoning and agent model from Z.AI, released under the permissive MIT license.

Parameters744 billion
Minimum GPU memory892.8 GB
Cheapest working build12x NVIDIA H100 SXM ($384,000)
Full breakdown
MIT

DeepSeek V4 Pro

DeepSeek

DeepSeek's largest open-weight model, built for top-tier agentic and coding performance, released under the permissive MIT license.

Parameters1,600 billion
Minimum GPU memory1920.0 GB
Cheapest working build24x NVIDIA H100 SXM ($768,000)
Full breakdown
OpenMDW

Nemotron 3 Ultra

NVIDIA

NVIDIA's flagship open-weight model, tuned for enterprise agent workloads, released under NVIDIA's Open Model Development Warranty license.

Parameters550 billion
Minimum GPU memory660.0 GB
Cheapest working build9x NVIDIA H100 SXM ($288,000)
Full breakdown