Holo3.1: Fast & Local Computer Use Agents

Hugging Face has announced the release of Holo3.1, an upgraded family of computer-use models designed to address production challenges across diverse environments, agent frameworks, and deployment targets. Following the initial adoption of Holo3 in browser and desktop workflows, Holo3.1 expands into mobile environments, yielding performance gains on benchmarks such as AndroidWorld where the 35B-A3B model score improved from 67% to 79.3%. Built on the Qwen family, Holo3.1 introduces native support for function-calling protocols alongside structured JSON outputs to better support third-party agent stacks. To enable cost-effective, private, and local inference on end-user devices, the release spans new model sizes including 0.8B, 4B, 9B, and 35B-A3B. Additionally, Hugging Face is shipping its first quantized weights, including FP8, Q4 GGUF, and NVFP4 checkpoints. Using NVIDIA's Model Optimizer, NVFP4 checkpoints achieve fast local inference with minimal performance degradation compared to full-precision BF16 checkpoints. On DGX Spark, NVFP4 W4A16 delivers 1.41 times the token throughput of FP8 and 1.74 times that of BF16. Combined with agent harness optimizations, NVFP4 quantization achieves a compound approximately two times end-to-end speedup, reducing average step time from 6.8 seconds to 3.3 seconds. Q4 GGUF checkpoints are also provided for local deployment on consumer hardware such as Apple Silicon, allowing agents to operate fully locally and privately on Windows or Mac machines.