Mission Control | field reports from the savage model lab
Mission Control
This is the editorial desk for Yeti Claw benchmark stories, render sweeps, latency curves, galleries, and the hard data behind what the fleet can really do under load.
Latest dispatch
Featured Dispatch | July 27, 2026
The new committee benchmark separates live-stack truth from isolation-lab truth on Spark: Nemotron keeps the default public lane, GLM stabilizes in the isolated Ollama window, Kimi remains the slower specialty path, and the co-resident GLM attempt failed with CUDA out-of-memory while the TRT services were resident.
Economics desk
The routing and product-behavior view that ties the economics, capacity, and image findings together.
The fleet-level comparison between published cloud GPU rates and the real hardware already running Yeti Claw.
The CPU-backed ESXi lanes now have a benchmark-backed place in the cost story as low-capital, queue-friendly text workers.
Capacity lab
Public-path capacity test for the dedicated BeastMode lanes after the static-IP cutover and selector publish.
Controlled simultaneous-conversation study showing where the interactive comfort band ends on the two primary text surfaces.
Media lab
Twenty DGX Spark cycles combining the younger influencer nightlife prompt with explicit hand and finger anatomy controls, plus Osyrus vitals and downloadable study artifacts.
Ten Spark video models ran the same horror prompt. Three clips completed, seven failures mapped to runtime, memory, or adapter issues.
The full rollout record for the 10-model Gojira video stack, from cold-start timeout through the first validated hosted Wan clip.
Eight of ten open-source generators completed under burst load, with vitals, failure notes, and gallery outputs archived.