Tencent launched Hy4 on November 4, 2026: a 770-billion-parameter model (about 49 billion activated per query) that understands images and operates software through command-line and graphical interfaces. The weights are not out yet — Tencent says they arrive in the coming days — so today Hy4 is an API product, not a downloadable model. That is fine. You can still evaluate it properly, if you evaluate the right things.
Here is the developer's checklist, in the order that saves you the most pain.
- Separate the hosted product from the future weights
Do not architect around a download that does not exist yet. Record the exact model identifier, the region, and the endpoint you tested — TokenHub, Yuanbao, CodeBuddy, WorkBuddy, or a third-party integration — because computer-use availability and pricing can differ per endpoint. When the weights land, re-run every test against the local build. If the numbers move, the endpoint was doing work the weights do not explain, and you want to know that before you ship.
- Price the trace, not the answer
Tencent quotes roughly $0.90 per million input tokens and $2.70 per million output tokens. A computer-use workflow is not one answer; it is a trace: screenshots or screen descriptions in context, a plan, tool calls, observations, retries, and a final confirmation. Output tokens and long contexts dominate that bill. Before any pilot, instrument your harness to log tokens per completed task, not per call, and compute cost per successful task including failed attempts. A model that looks cheap per token and expensive per completed task is expensive.
- Test on your task, with a bounded account
Vendor demos use clean screens and forgiving tasks. Your evaluation should use the messiest real workflow you actually want to automate, on a test account with no destructive permissions: no payments, no deletions, no external submissions without human approval. Measure four things: task completion rate, cost per completed task, recovery rate after a wrong step (does the agent notice, stop, and undo?), and untrusted-content handling (does it follow instructions embedded in a webpage or a pasted document?). That last one is a security test, not a capability test — an agent that obeys screen content is an agent a stranger can steer.
- Watch the failure modes that are specific to GUI agents
Text models fail with wrong words. GUI agents fail with wrong clicks: selecting the wrong row, typing into the wrong field, confirming a dialog it did not read. Log every action with a screenshot or DOM snapshot before and after, so failures are debuggable rather than mysterious. Set step limits and time limits per task; a long-horizon agent that loops is a billing event as well as a bug. And keep a kill switch: one command that revokes the agent's session immediately, tested before the pilot, not during an incident.
- Compare under identical conditions
Run the same tasks, same harness, same scoring against your current best alternative — another computer-use model, a scripted RPA flow, or a human baseline. Report completion, cost, and error rate side by side. A leaderboard number from a vendor, without the task list and the harness, is not a comparison. Your harness is.
- Do not skip the licence and system-card check
When the weights and system card arrive, read them before you fine-tune or self-host: licence terms for commercial use, disclosed training-data scope, safety evaluations, and stated limitations. A permissive announcement and a restrictive licence are different products. If the card is thin, treat the model as unproven for regulated or customer-data workloads until independent evaluations fill the gaps.
- Record the trace, not just the score
A pass/fail score without the trace is undebuggable. Store the model identifier, the prompt, each tool call, each screen observation, and the cost for every run. When a task fails in production — and one will — that trace is the difference between a one-line fix and a week of guessing. It is also the evidence you show a customer who asks why the agent did what it did.
The bottom line
Hy4's combination — vision, GUI and CLI operation, sparse activation at frontier scale — makes it one of the most interesting computer-use releases of the year. But interesting is not the same as ready. The disciplined move this week is a small, instrumented pilot: one workflow, bounded permissions, full traces, cost per completed task. If Hy4 wins that pilot, you have evidence. If it does not, you have saved yourself a production incident and a bill. Either way, you will be ready the day the weights actually land — which is when the real evaluation begins.
Sources: Tencent Hy4 launch reporting, November 4, 2026; Kimkj daily briefing, November 5, 2026 (770B total parameters, ~49B activated per query). Independent evaluations pending. Author: Shivam Kumar, founder of VisionQuantech.
