BLOCKWAY · ON-DEVICE MODEL · HONG KONG
Sees the screen. Does the work. A 1.2B vision-language-action model that grounds every click to the pixel across macOS, Windows and Android — with no pixel ever leaving the device.
Commercial licence · private deployment · part of the Blockway model ecosystem
350MB
INT4 build — loads on phone NPUs
597M
active parameters (1.2B total)
64K
context, constant memory over long sessions
100%
offline — zero pixels to the cloud
Screens are the one interface every legacy system already has. Argus closes the loop on them — observe, ground, act, verify — without an API, and without the cloud.
A dedicated grounding head maps the instruction straight to a 36×36 cell grid over the live screen — any resolution, any aspect ratio — with a second zoomed pass to refine small targets like menu-bar icons and system trays.
A built-in planner decomposes a natural-language task into steps, re-observes the screen after every action, and carries multi-window, multi-app workflows through to the end.
Each action ships with its reasoning trace and coordinates, recorded for replay. For finance, healthcare and government, that audit trail is the precondition for automation — not a nice-to-have.
Trained specifically on Chinese operating systems, websites and business software. Mixed Chinese-English interfaces — the reality of Hong Kong enterprise desktops — are a first-class citizen, not an afterthought.
A hybrid Mamba-2 backbone (22 state-space + 3 attention layers) keeps memory growth far below a standard attention stack as the session grows. Argus can watch an entire evening of workflow on a phone without the quadratic blow-up a Transformer would hit.
About 350 MB at INT4 — small enough for the Apple Neural Engine, Snapdragon X NPUs and Jetson Orin Nano, and served on Apple-silicon Macs via MLX today.
ONE MODEL · THREE PLATFORMS
Element-grounding hit rate — the predicted click landing inside the target element — for the latest unified Argus against its previous base, on public ScreenSpot and on internal held-out sets collected from real macOS, Windows and Android screens.
ScreenSpot (public)+2.8
1,272 cases · two-pass grounding
macOS desktop+29.2
internal held-out · 240 cases
Windows desktop+11.0
internal held-out · 200 cases
Android+11.2
internal held-out · 80 cases
ScreenSpot, macOS and Windows use two-pass zoom grounding; Android is single-pass. Internal sets are held out from training by screen. Figures as of August 2026; evaluation method ships with the enterprise technical report.
THE LANDSCAPE
Raw accuracy was never the battleground for on-device agents — accuracy × size × privacy × deployability is. Argus is the only model below with a phone-NPU-sized footprint, full offline operation and an auditable trace.
ScreenSpot accuracy vs model size
Competitor figures are the ScreenSpot averages published in each model's own technical report — UI-TARS, ShowUI, SeeClick, CogAgent; evaluation details differ across papers, so treat cross-model numbers as indicative. “Phone NPU” means the INT4 build loads on current handset neural processors at interactive latency.
THE BLOCKWAY MODEL ECOSYSTEM
Blockway ships models as a working ecosystem: Agens Pilot reasons and converses; Argus watches the screen and operates it. Together they power Claway's AI employees on hardware the customer owns.
Where Argus goes to work
Claway's AI employees take on whole roles on a company's own machines. Argus is the part that operates the software humans use — the 40-year-old mainframe, the desktop ERP, the government portal — clicking and typing where no API exists.
Visit claway.ioThe reasoning sibling
Blockway's Cantonese-first open model handles the conversation, the judgement and the long-context reasoning; Argus turns the resulting plan into on-screen action. Two models, one workforce.
Explore Agens PilotA hybrid Mamba-2 + attention backbone (597M active parameters), an any-resolution vision tower and a single-pass grounding head — packaged for each class of hardware. Delivered under commercial licence to enterprise partners; weights are not publicly released.
≈ 350 MB
≈ 1.6 GB
≈ 3.4 GB
WHERE THE CLOUD CANNOT GO
Forty-year-old core systems, internal software with a UI but no API, regulated data that cannot cross a border — exactly the places cloud agents are excluded from are where Argus works.
Law firm
Photograph a client's Chinese contract on a phone. Argus identifies the contract type, locates twenty-plus key clauses and drafts response points — on the device, in seconds.
Zero uploads. Confidentiality obligations kept intact.
Bank counter
Facing a decades-old mainframe UI, staff highlight the customer's message and say “open a new account”. Argus navigates five screens and fills thirty fields; a human only reviews and approves.
No core-system changes. Counter throughput up from day one.
Insurance claims
Shoot the scene on site. Argus assesses the damaged areas, matches policy clauses, estimates the payout range and drafts the claim form — entirely on the device.
Customer privacy never leaves the handset; claim cycles shrink.
Argus ships to enterprise partners under commercial licence, with domain fine-tuning and fully private deployment. We are onboarding design partners in finance, legal, insurance and manufacturing.
