Zen 6 Max
3.7 trillion parameters across a routed fusion. Each request runs on the specialist Enso picks, so it behaves as one model.
One name
You call Zen 6 Max. You never choose among the models behind it.
The right specialist, per request
Enso reads each request and runs it on the specialist built for that kind of work, transparently.
Long context
A long context, for large codebases and long documents.
Tools and code
Tool calls and code, built in.
Enterprise deployment
On your own premises
What a rack for Zen 6 Max is made of.
We recommend at least four 8-GPU NVIDIA Blackwell servers, B300 or B200, for Zen 6 Max on your own premises. Routing cuts the average inference work a request needs. It does not cut the memory needed to keep every specialist ready.
DGX B300
DGX B200
Four B300 servers hold 9,216 GB of GPU memory; four B200 servers hold 5,760 GB.
Planning range
Qualification. Four 8-GPU servers is where a qualification starts.
Isolated production pools. One pool for each specialist, so a busy class of work never takes memory from another.
Redundant pools. Two independent replicas of those pools across failure domains.
Around the servers
NVLink and NVSwitch inside each server.
400 or 800 Gb/s RDMA wherever a specialist spans servers.
About 31 TB of NVMe in each server (8 x 3.84 TB).
Separate sandbox workers for coding agents.
A high-availability control cluster.
Final capacity and performance commitments come from qualifying your workload: context length, concurrency, routing mix, latency, and failover.