H Company, NVIDIA launch Holotron-12B for agents

This looks like a practical shift from “smart demo model” to “deployable agent policy model,” with throughput as the main differentiator rather than raw parameter count.

• –The architecture focus (hybrid SSM-attention) is aimed at long-context, multi-image workloads where transformer KV-cache costs usually bottleneck agent systems.
• –Reported scaling to about 8.9k tokens/s at concurrency 100 (vs ~5.1k for Holo2-8B in the same setup) is the headline claim for teams running many parallel agent trajectories.
• –WebVoyager improvement from 35.1% to 80.5% versus the Nemotron base suggests post-training data quality, not just architecture, is doing heavy lifting.
• –The NVIDIA Open Model License makes it accessible for experimentation, but production users will still want independent replication of throughput and task-completion metrics.