H Company Ships Holo3.1 with Local Inference, Mobile Support, and 79.3% AndroidWorld Score

H Company released Holo3.1, an updated family of computer-use agent models designed for cross-environment deployment and local inference. The release includes four model sizes (0.8B, 4B, 9B, and 35B-A3B parameters) and marks H Company's first shipment of quantized checkpoints for on-device execution.

Performance Gains Across Mobile and Desktop

Holo3.1-35B-A3B achieves 79.3% on AndroidWorld, a 12.3 percentage point improvement over Holo3's 67%. Smaller variants also show gains: the 4B and 9B models improve from 58% to 72% on the same benchmark.

Built on the Qwen model family, Holo3.1 adds native function-calling support alongside the structured JSON outputs available in Holo3. According to H Company, function-calling and native execution now achieve near-parity performance across OSWorld and internal benchmarks covering e-commerce and business software workflows.

The company reports a 25% improvement over Holo3 when evaluated inside its Holotab product harness, though absolute scores were not disclosed.

Quantized Checkpoints for Local Deployment

H Company ships FP8, Q4 GGUF, and NVFP4 quantized weights for the 35B-A3B model. The NVFP4 checkpoint uses NVIDIA's Model Optimizer in a W4A16 configuration.

On DGX Spark hardware, NVFP4 W4A16 delivers 1.41× the token throughput of FP8 and 1.74× that of BF16, according to H Company's benchmarks. FP8 and NVFP4 match OSWorld scores, trailing the BF16 checkpoint by approximately two points.

End-to-end agent step time drops from 6.8 seconds to 3.3 seconds on Spark with NVFP4 quantization and agent harness optimizations, representing a 2× compound speedup. H Company states these optimizations will ship in an upcoming desktop agent harness.

The Q4 GGUF checkpoints target consumer hardware including Apple Silicon, enabling fully local execution where both the agent and model run on the same machine or local network.

Model Specifications

The Holo3.1 family includes:

• Holo3.1-0.8B: Ultra-lightweight local agents
• Holo3.1-4B: Cost-efficient deployment
• Holo3.1-9B: Balanced performance and latency
• Holo3.1-35B-A3B: State-of-the-art performance

Pricing, context window size, and training data cutoff date were not disclosed. Models are available through H Company's API and Hugging Face.

What This Means

Holo3.1's quantized checkpoints address a practical deployment constraint: running computer-use agents on local hardware without cloud dependencies. The 2× speedup on DGX Spark and sub-4-second step times make interactive agent workflows more viable. The 12-point AndroidWorld improvement suggests H Company expanded its training or fine-tuning data to cover mobile interfaces more thoroughly. However, without disclosed benchmark scores on standard desktop tasks like OSWorld or pricing details, it's unclear whether the quantization tradeoffs favor local deployment over cloud inference for most use cases.