Domain post-training, frontier-class local models, and a multi-stream agent runtime in one box. Real-time professional reasoning, with data that never leaves your domain. Desktop-class hardware: power and ethernet, and it stands watch — no machine room required.
Wherever there are people, business and private data, intelligence should not live in someone else's rack.
Official model-card public benchmarks: several capabilities exceed current online flagships; the overall profile sits in the frontier range — local deployment no longer means compromise.
| Benchmark | AI-hub built-in model | Claude Opus 4.8 | GPT-5.6 Sol | DeepSeek V4 Flash |
|---|---|---|---|---|
| SWE-bench Pro · real code repair | 61.7 | 69.2 | 64.6 | 56.0 |
| OSWorld-Verified · desktop ops | 84.3 | 83.4 | 83.2 | — |
| IFBench · instruction following | 79.5 | 62.2 | 72.7 | — |
| GPQA Diamond · science reasoning | 89.2 | 93.6 | 94.1 | 90.8 |
| Terminal-Bench 2.1 · terminal coding | 73.0 | 74.6 | 88.8 | 82.7 |
Desktop operation and instruction-following lead both overseas flagships (OSWorld 84.3 vs 83.4 / 83.2; IFBench 79.5 vs 62.2 / 72.7); the GPQA science gap with each flagship is under 5 points; LiveCodeBench v6 90.3, OmniDocBench 1.5 91.1 (most flagships publish no comparable figures) — the overall profile sits in the current frontier range.
Sources: official model cards & system cards of each model (vendor-reported, August 2026 snapshot); DeepSeek V4 Flash from its official card and third-party evaluation compendia.
Operators are built for the model, the model is tuned for the business, and the runtime hands every terminal's request to the right model. The layers mesh, and it works on arrival at the site.
Domain protocols, playbooks, tool interfaces and expert knowledge are baked into the device at model-customization time. Ready at power-on, shared across sessions, zero switching cost. Co-innovation across silicon operators, model restructuring and agent interaction — all you feel is the speed.
A frontier-class local model, post-trained on industry and specialist knowledge. 50–100 chars/second of sustained output — roughly ten times human thought speed. Complex reasoning, long-document understanding, anomaly assessment, with domain-expert-grade decision proposals.
Sixteen independent agents run in parallel on one machine, enrolled against the org chart with fully isolated contexts. Sensing, assessment, planning and instruction each do their own job; a single-instance failure never spreads.
Intent recognition, task decomposition, local API and equipment invocation, result validation, plan assembly — one chain completed inside the inference pipeline.
Business documents and product FAQs hot-update; retrieved results merge into context automatically, no restarts.
A lightweight model acts as gateway, judging request difficulty — simple intents answered instantly, complex reasoning handed to the deep tier. Compute spent where it counts.
When the internet drops, core reasoning and agent services keep running. The outside network is an update channel, not a lifeline.
Every instruction and every disposition is traceable, replayable and auditable — the action chain closes completely.
The same engineering path as autonomous driving: a four-step closed loop of learn, forge, connect, decide. Every use makes the system stronger.
Literature, protocols, wargaming material, case files and equipment manuals — all domain material becomes training data. Decades of accumulated expertise becomes machine-usable capability for the first time.
Industry post-training and compression on a fully self-reliant base model, delivered inside the on-site device, producing national-expert-level decision proposals.
Cameras, lidar, satellite feeds and gauge data converge; uncrewed vehicles, drones, automatic devices and valves come under unified control. The machine sees, and it can reach.
Machine-generated response plans go to execution only after the responsible person reviews and approves; execution data flows back, effectiveness is assessed in real time, and the model keeps getting stronger.
A desktop box in the duty room or command center — a normal office environment is enough, no machine room. Cameras, radar and gauges plug in; uncrewed equipment comes under command.
The hub exposes standardized interfaces. Sensing and execution devices only need two things: reach the LAN, and understand the commands.
from k3box import AgentClient client = AgentClient("http://hub.topobius.local:8080") session = client.session(agent="site-command", terminal="coa-01") reply = session.chat("Assess the situation; propose three options") # → the machine proposes, the person decides
Python / Node.js / Java SDKs, HTTP-JSON for existing systems, WebSocket streaming output. Every instruction and every disposition is traceable and auditable.
| Topobius Gradientron | Per-terminal deployment | Public-cloud API | |
|---|---|---|---|
| Intelligence level | Frontier-class | Compromised | Frontier-class |
| Hardware | 1 device | One per person, device churn | None |
| Data boundary | Stays in-domain | Stays on terminal | Leaves to cloud |
| Network dependence | Runs offline | Runs offline | Hard dependency |
| Latency | < 50 ms (LAN) | < 50 ms (local) | 100–500 ms (WAN) |
| Ops | 1 device, centrally managed | Walks out with the person | Vendor-dependent |
| Running cost | ≈ ¥50/mo electricity | ≈ ¥80/mo electricity | Per-token, volatile |
| Compliance risk | Low | Low | Medium-high (data export) |
CNC-milled all-aluminum unibody, at home in the duty room, command center or office. Key component figures follow the formal delivery list.
FLEET DEPLOYMENT · IDENTICAL BUILD
THREE-QUARTER VIEW · WARM DESK
HIGH ANGLE · DESKTOP PRESENCEBook an on-site demo: bring one of your existing terminals, plug it in, and watch the first-token latency, the concurrency, the data flow.