Holo 3.1 35B A3B on live screen streams

What Holo 3.1 35B A3B is good at

Holo 3.1 was trained specifically for computer-use agents: it locates buttons, fields, and menus on a screenshot, understands what each element does, and decides the next action in a multi-step workflow. That specialization shows up on agent benchmarks, where the 35B A3B variant reaches 79.3% on AndroidWorld and extends the earlier Holo3 line beyond browser and desktop automation into mobile environments.

The model also ships with native function-calling support, so it plugs directly into agent frameworks that express clicks, keystrokes, and scrolls as tool calls. It is the largest member of the Holo 3.1 family, which spans 0.8B to 35B-A3B sizes, and the weights are released under Apache 2.0, which permits commercial use without a restrictive license gate.

• UI grounding: precise localization of on-screen elements from screenshots
• Multi-step action planning across web, desktop, and mobile interfaces
• Native function calling for direct integration with agent frameworks

Running Holo 3.1 35B A3B on Overshoot

Create a Stream, publish a screen share or camera track over WebRTC through LiveKit, then send a chat-completions request whose image_url is an ovs:// reference. Overshoot retains 600 seconds of frame history, so a request can anchor the latest frame, an exact timestamp, or a bounded recent segment, which suits agents that need to confirm what just changed on screen.

Because the model is open-weight, Overshoot serves it on the hosted fast path under the api id Hcompany/Holo-3.1-35B-A3B-FP8, an FP8 checkpoint tuned for low-latency inference. Responses stream token by token over SSE with typical answers in about 200ms, and reusing thread_id across queries against the same stream keeps hitting the prompt cache, which matters when an agent loop calls the model once per action.

Sparse MoE efficiency for agent loops

Computer-use agents are unusually latency-sensitive because a single task can require dozens of model calls: observe the screen, pick an action, act, observe again. Holo 3.1's sparse mixture-of-experts design activates only about 3B parameters per token, so each call carries near-small-model serving cost while the full 35B parameter pool preserves flagship-level perception and planning quality.

The 262K-token context window compounds that advantage. Long automation sessions can carry conversation history, prior screenshots, and function-calling schemas in a single thread without aggressive truncation, so the agent keeps full task context from the first click to the last.

How it compares to other agent models

Against dense computer-use models such as UI-TARS 72B, Holo 3.1 35B A3B trades raw parameter count for activation efficiency: far fewer active parameters per token, with grounding and action-planning accuracy that holds up on agent benchmarks. Compared with generalist vision models like Qwen3-VL 32B or GLM-4.5V, it gives up some breadth on open-ended visual question answering in exchange for sharper UI understanding.

Within its own family, the smaller 0.8B, 4B, and 9B Holo 3.1 checkpoints target on-device and edge deployment. The 35B A3B variant is the accuracy pick, and it is the one served on Overshoot's hosted path for live streams.

Frequently asked questions

Is Holo 3.1 35B A3B available on Overshoot?

Yes. Holo 3.1 35B A3B is live in the Overshoot catalog as a hosted model with ready status, served under the api id Hcompany/Holo-3.1-35B-A3B-FP8. Requests use the same OpenAI-compatible chat-completions shape as every other model on the platform.

Can Holo 3.1 35B A3B analyze live video?

Yes. Through Overshoot, Holo 3.1 35B A3B answers questions about a live WebRTC stream, including screen shares, using ovs:// frame references inside a standard chat-completions request. Answers stream back over SSE, typically in about 200ms on the hosted fast path.

Is Holo 3.1 35B A3B open source?

Weights are released under Apache 2.0 on Hugging Face, alongside quantized FP8, NVFP4, and GGUF builds. That open availability is what lets Overshoot serve it on the low-latency hosted path rather than as a passthrough.

What is Holo 3.1 35B A3B best used for?

It is built for computer-use agents: grounding UI elements on a live screen share, planning multi-step actions across web, desktop, and mobile interfaces, and driving automation loops through native function calling. General visual question answering works, but UI-centric tasks are where it stands out.

Related models

• UI-TARS 72B
• GLM-4.5V
• Qwen3-VL 32B
• Kimi-VL A3B
• Holo3 35B A3B