The Gap Between Benchmarks and Shipping
When H Company released Holo3.1, they didn't lead with a new state-of-the-art benchmark score. Instead, they led with deployment targets: FP8, Q4 GGUF, NVFP4 quantization. Function-calling protocols. Mobile environments. Agent harnesses.
This is what maturity looks like. The computer-use agent space is moving from "can we do this?" to "can we ship this?" And shipping means dealing with the messy realities of production: cost, latency, privacy, cross-platform support, and integration with existing agent frameworks.
Holo3.1 is H Company's answer to those constraints. It's not just a better model—it's a model family designed for the real-world trade-offs teams actually face when deploying agents.
Four Sizes, Three Quantization Formats
The release spans four model sizes: 0.8B, 4B, 9B, and 35B-A3B. The naming tells you the intent: ultra-lightweight local agents, cost-efficient deployment, balanced performance, and state-of-the-art capability, respectively.
What's more interesting is the quantization story. This is H Company's first release to ship optimized weights in FP8, Q4 GGUF, and NVFP4 formats. They used NVIDIA's Model Optimizer in W4A16 configuration for the latter.
The performance retention is impressive. On OSWorld, FP8 and NVFP4 match each other and sit only about two points below the full-precision BF16 checkpoint. That's minimal degradation for what turns out to be substantial throughput gains.
On DGX Spark, NVFP4 W4A16 delivers 1.41× the throughput of FP8 and 1.74× that of BF16. The agent harness optimizations they developed with NVIDIA compound this further: average step time drops from 6.8 seconds to 3.3 seconds—a ~2× end-to-end speedup over the FP8 baseline.
That's the difference between a research artifact and a shipping product.
Mobile and Cross-Harness Reality Checks
The other major theme in this release is robustness across execution environments. H Company observed the classic transfer problem: strong performance in one setting doesn't guarantee it in another. Mobile devices, alternative agent frameworks, different execution harnesses—each introduces its own distribution shift.
On AndroidWorld, the 35B-A3B model improves from 67% to 79.3%. The smaller 4B and 9B variants jump from 58% to 72%. That's a 12-14 percentage point gain across the board.
On the agent framework side, Holo3.1 adds native support for function-calling protocols alongside the structured JSON outputs Holo3 already provided. Across OSWorld and H Company's internal benchmark suite covering e-commerce, business software, and collaboration workflows, function-calling and native execution now achieve near-parity performance.
They also report a 25%+ improvement over Holo3 when evaluated inside their Holotab product harness. That last detail is revealing—it suggests the delta between benchmark performance and in-product performance was larger than they wanted.
Local Inference on Consumer Hardware
Here's where it gets geeky: H Company is targeting fully local deployment of computer-use agents on consumer hardware. The agent runs locally on Windows or Mac, and the model either runs on the same machine or on a DGX Spark on the same network. Either way, nothing leaves the user's network.
They include reference numbers for Apple Silicon. The Q4 GGUF checkpoints are explicitly aimed at this use case.
This matters more than it might seem. Computer-use agents see your screen, click your buttons, and interact with your applications. Privacy sensitivity is extreme. For enterprise workflows especially, keeping everything on-premises or on a local device is often non-negotiable.
The latency story matters too. Computer-use agents are step-by-step systems. If each step takes 6-8 seconds, the user experience degrades fast. Getting to 3.3 seconds average step time on local hardware changes what kinds of workflows feel responsive enough to be useful.
The Cost-Performance Frontier
H Company published a Pareto chart showing performance versus cost across the Holo3.1 and Qwen 3.5 families. The performance metric averages four H Corporate benchmarks (so each family is equally weighted), then takes the mean across OSWorld, AndroidWorld, H Corporate, ScreenSpot-Pro, and OSWorld-G.
The visual takeaway: the smaller Holo3.1 models punch above their weight. The 4B and 9B variants occupy attractive points on the curve—cheaper than the 35B model but still competitive with much larger alternatives on aggregate performance.
This is the practical side of computer-use agents. Not every task needs state-of-the-art capability. Many workflows are cost-sensitive, latency-sensitive, or privacy-sensitive. Having a model family that spans 0.8B to 35B gives teams actual deployment options.
What's Still Missing
H Company doesn't publish absolute OSWorld or AndroidWorld scores for the smaller models in the blog post, just the improvement deltas. That makes it harder to assess whether the 4B or 9B variants hit the capability floor needed for real-world tasks, or whether they're still mostly experimental.
They also don't detail what the Holotab product harness actually is, beyond it being an internal benchmark. That's fine—internal product metrics are proprietary—but it makes the 25% improvement claim harder to contextualize. Improvement over what baseline? Doing what tasks?
The quantized weights are currently only available for the 35B-A3B model. If the smaller models are meant for local inference, you'd expect Q4 GGUF checkpoints for those too. Maybe they're coming.
The Maturity Signal
The most important thing about this release isn't any single benchmark number. It's the focus on deployment targets, quantization formats, cross-platform support, and agent framework compatibility.
Computer-use agents are moving out of the eval harness and into production. That means dealing with the constraints of real systems: cost budgets, latency requirements, privacy policies, integration complexity, and the sheer diversity of platforms people actually use.
Holo3.1 is H Company saying: we're ready to ship.
You can try the models via the Holo Models API or grab them directly from the Hugging Face collection.