H Company · v3.1 · depuis 2. Juni 2026 · 6× · vu le 29 juin 2026
5
Momentum
Holo3.1 is a family of vision-language models for computer-use agents released by H Company (Paris) in June 2026, built on the Qwen architecture. The models come in four sizes (0.8B, 4B, 9B, 35B-A3B) and cover web, desktop, and mobile environments. New quantized checkpoints (FP8, Q4 GGUF, NVFP4) enable, for the first time, fully local execution on Windows, Mac, or DGX Spark hardware, alongside native support for function-calling protocols for integration into agent frameworks. Open weights are released under the Apache 2.0 license on Hugging Face, while cloud API access via the Holo Models API is billed on a usage basis.
Historique du momentum
19.05.17.08.
Fonctionnalités
| Autonomy Level | Autonomous computer-use agent: analyzes screenshots and independently performs actions such as clicks, keystrokes, and scrolling |
| Channels | Web, desktop, and mobile environments (browser automation, business software, Android devices) |
| License | Apache 2.0 (open weights on Hugging Face) |
| Modalities | Multimodal: text and image input (screenshots) for visual UI interpretation and action planning |
| Orchestrated AI Models | Qwen 3.5 family (base models); Fine-tuned VLMs: Holo3.1-0.8B, Holo3.1-4B, Holo3.1-9B, Holo3.1-35B-A3B (35B total / 3B active, MoE); Quantizations: BF16, FP8, NVFP4 (W4A16), Q4 GGUF |
| Platform | Runs locally on Windows or Mac machines, with optional model execution on Apple Silicon or a DGX Spark on the same network; cloud API also available |
| Price | Open-weight models free (Apache 2.0, self-hostable); cloud API has free rate-limited tier, then usage-based billing per million tokens |
| Release Date | June 2, 2026 |
| Tool/MCP Integrations | Native function-calling support in addition to structured JSON outputs; OpenAI-compatible Models API for third-party agent frameworks |