Hosted on OvershootHolo3 35B A3B on live video streams
Holo3 35B A3B · H Company · Computer-use · Open-weight
Holo3 35B A3B is H Company's open-weight computer-use vision-language model, a sparse Mixture-of-Experts design with 35B total parameters and only 3B active per token. Released in March 2026 under Apache 2.0, it posted a state-of-the-art 77.8% on OSWorld-Verified while costing a fraction of what larger proprietary agents charge per step. On Overshoot it is listed on the hosted fast path under the api id Hcompany/Holo3-35B-A3B, though the catalog entry currently reports unavailable status, so check GET /v1beta/models before routing traffic.
Holo3 35B A3B runs on Overshoot's hosted fast path as Hcompany/Holo3-35B-A3B (status: unavailable, verified against the live model catalog on 2026-07-14).
- Developer
- H Company
- Parameters
- 35B total, 3B active (MoE)
- Base model
- Qwen3.5-35B-A3B
- Context window
- 262K tokens
- License
- Apache 2.0
- Released
- Mar 2026
- Inputs
- Text, images, video frames
- OSWorld-Verified
- 77.8%
- Overshoot availability
- Hosted fast pathas of 2026-07-14
What Holo3 35B A3B is good at
Holo3 35B A3B is built for GUI agents: it interprets visual interfaces across web, desktop, and mobile environments, reasons over on-screen content, and outputs precise actions such as clicks, form fills, spreadsheet edits, and browser navigation. Its 77.8% score on OSWorld-Verified led the benchmark at release with only 3B active parameters per inference.
Beyond end-to-end task execution, the model delivers top-tier UI localization and grounding on ScreenSpot-Pro and OSWorld-G, meaning it can point at the exact element an agent needs to act on rather than just describing the screen. H Company also validated it on an internal benchmark of 486 multi-step business tasks spanning e-commerce, office software, and multi-app workflows.
It retains the strong web navigation behavior of the earlier Holo2 line while extending it into realistic desktop and enterprise environments, which makes it a natural policy model for automation that watches a live screen share.
- State-of-the-art computer-use execution on OSWorld-Verified
- Precise UI element grounding on ScreenSpot-Pro and OSWorld-G
- Multi-step business workflows across web, desktop, and mobile
Running Holo3 35B A3B on Overshoot
Create a Stream, publish a screen share or camera over WebRTC through LiveKit, then send an OpenAI-compatible 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 agent loops that need to re-check the screen after each action.
Holo3 35B A3B is catalogued on Overshoot's hosted fast path under the api id Hcompany/Holo3-35B-A3B. Hosted models stream token by token over SSE and typically answer in about 200ms, a budget that matters for computer-use agents issuing many observe-act cycles per task.
Note that the model currently reports unavailable status in the live catalog. It remains listed, but you should confirm its state with GET /v1beta/models before pointing production traffic at it, and fall back to another agentic model if it has not returned.
Why the sparse MoE design matters
The A3B suffix describes the architecture: 35B total parameters organized as a sparse Mixture-of-Experts, with roughly 3B active on any forward pass. That keeps per-token compute close to a small model while preserving the capacity of a much larger one, which is why Holo3 undercuts dense computer-use models on inference cost at comparable accuracy.
Holo3 35B A3B is fine-tuned from Qwen3.5-35B-A3B and inherits its 262K-token context window, enough to carry long action histories, accessibility trees, and many sampled frames in a single agent thread. The Apache 2.0 license also permits self-hosting and fine-tuning without restrictive terms.
Holo3 versus Holo3.1 and other computer-use models
H Company followed up with Holo3.1 35B A3B in June 2026, focused on faster local execution, and began deprecating the original Holo3 endpoint upstream in mid-June. Teams standardized on Holo3's exact action behavior may still prefer it, but new projects should evaluate the 3.1 line as the successor.
Compared with UI-TARS 72B, the other dedicated computer-use model in the Overshoot catalog, Holo3 activates far fewer parameters per step, trading raw dense capacity for much lower latency and cost across long agent trajectories. General visual reasoning models like QVQ 72B remain better fits when the task is analysis rather than acting on a UI.
Frequently asked questions
Is Holo3 35B A3B available on Overshoot?
It is in the live Overshoot catalog as a hosted model under the api id Hcompany/Holo3-35B-A3B, but it currently reports unavailable status. Check GET /v1beta/models for its current state before sending requests, and have a fallback agentic model configured.
Can Holo3 35B A3B analyze live video?
Yes. Through Overshoot, it reads frames from a live WebRTC stream, typically a screen share, referenced with an ovs:// URL inside a standard chat-completions request. Responses stream back over SSE, and the hosted fast path targets roughly 200ms answers when the model is live.
Is Holo3 35B A3B open source?
Yes. H Company released the weights on Hugging Face under the Apache 2.0 license, which permits commercial use, self-hosting, and fine-tuning. That open availability is what qualifies it for Overshoot's hosted serving path.
What does A3B mean in the model name?
It denotes the sparse Mixture-of-Experts layout: 35B total parameters with about 3B active per token. The model routes each step through a small subset of experts, so it runs at near-small-model speed and cost while keeping large-model capability on GUI tasks.