Holotron-12B - High Throughput Computer Use Agent
Hugging Face
H Company released Holotron-12B, an open model built to control computers - clicking, navigating, filling forms - way faster than before. It's twice as fast as their last model at the same job, which matters if you want AI agents doing real work at scale.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
H Company just put out Holotron-12B, and the pitch is refreshingly narrow: this isn't another do-everything vision model, it's a policy engine for agents that have to look at a screen and act on it, over and over, without falling over under load. It's built on top of NVIDIA's Nemotron-Nano-2 VL, then fine-tuned by H Company on their own screen-navigation and UI-grounding data, roughly 14 billion tokens worth.
The interesting part is architectural, not just behavioral. Nemotron uses a hybrid setup that mixes attention with state-space models, and SSMs sidestep the usual KV-cache bloat that chokes long-context inference in standard transformers. Instead of storing key-value pairs for every token at every layer, an SSM keeps a fixed-size state per layer regardless of how long the interaction history gets. For an agent juggling dozens of high-res screenshots across a long session, that's the difference between graceful scaling and hitting a wall.
And the numbers back it up. On the WebVoyager benchmark, running on a single H100 with vLLM 0.14.1, Holotron-12B pushed throughput to 8.9k tokens per second at 100 concurrent requests, while H Company's earlier Holo2-8B topped out around 5.1k tokens per second and plateaued well before that. More than double the throughput, from a model that's actually bigger. That's not a small footnote — it flips the usual assumption that bigger means slower.
Accuracy moved too, not just speed. Holotron-12B's WebVoyager score jumped from 35.1% on the bare Nemotron base to 80.5% after fine-tuning, and it edged out Holo2-8B on the same test. It also picked up solid gains on grounding benchmarks like OS-World-G, GroundUI, and WebClick — the tests that check whether a model can actually point at the right button rather than just describe it.
H Company is framing this as a stepping stone. NVIDIA announced Nemotron 3 Omni the same day, with a beefed-up SSM-attention mix and mixture-of-experts layers, and H Company says they're already planning to post-train that as the next Holotron. The stated goal is commercial-grade, high-throughput computer-use agents for enterprises, not just a research demo — which tells you where this whole line of models is headed next.
My take — AI-written commentary, not fact-checked reporting
What actually matters here isn't the benchmark score, it's the throughput chart — a 2x speedup at scale is the kind of unglamorous infrastructure win that determines whether agentic AI ships as a product or stays a cool demo burning GPU credits. Open-weighting this under an NVIDIA license while chasing enterprise deployment is the right call, and it's a sharper signal about where 2026 agent competition goes than any chatbot leaderboard.
Read more about this at: Hugging Face