跪拜 Guibai
← All articles
Agent · AI Programming · Tencent

Tencent's Hy3 Model and WorkBuddy Agent Run a Five-Case Gauntlet

By 沉默王二 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 295B MoE model that sustains nearly 500-step agent workflows and keeps data consistent across Excel, charts, and PPT changes the ceiling for unattended desktop automation. The two-week free trial lowers the barrier to stress-testing it on real office pipelines.

Summary

Hy3 is a 295B-parameter Mixture of Experts model with 21B active parameters and a 256K context window, now integrated into Tencent's WorkBuddy desktop agent. A five-case test suite exercises four capability dimensions: front-end planning, tool use, long-range execution, and complex reasoning. The front-end cases demand pixel-precise Neo-Brutalist design and an animated MoE routing visualization; the tool-use case cross-references GitHub, HuggingFace, and OpenRouter for data consistency; the long-range case reads a folder of Markdown files, produces an Excel report with charts, and generates a 10-page Swiss-style PPT with a self-check step. The reasoning case requires a full derangement-problem derivation, a million-run Monte Carlo simulation, and a convergence plot against the exact solution.

WorkBuddy completed each case with correct parameter extraction, data-consistent cross-tool output, and a simulation error within ±0.001 of the mathematical exact value. The agent also handled context compression mid-task during the heavy PPT generation run. Hy3 is available with a two-week free trial on WorkBuddy.

Takeaways
— Hy3 is a 295B total / 21B active MoE model with 192 routed experts (top-8 activated) plus one shared expert, supporting up to 256K context.
— Reasoning intensity is switchable across three levels: no_think, low, and high.
— WorkBuddy read a WeChat article URL, extracted benchmark figures (hallucination rate 12.5%→5.4%, task resolution 72%→90%, token efficiency gains up to 49%), and generated a poster with all numbers correct.
— A Neo-Brutalist single-page site was generated from a prompt specifying 4 hex colors, 3px solid borders, 6px hard shadows, mixed corner radii, inline SVG decorations, and hover interactions—all constraints were applied in one pass.
— An animated MoE routing visualization was built as a single HTML file with a token-flow animation, expert-lighting logic, and a speed slider, plus GIF frame-rate adjustment for WeChat compatibility.
— Cross-referencing GitHub, HuggingFace, and OpenRouter correctly identified that OpenRouter lacks the full Hy3 model and surfaced the 262K preview discrepancy.
— A long-range task read a folder of Markdown files, produced an Excel sheet with word counts and monthly stats, generated charts, and output a 10-page Swiss-style PPT—with a self-check confirming data consistency.
— Hy3 handled context compression mid-task during the heavy PPT generation run.
— A derangement problem was solved with full inclusion-exclusion derivation, a 1-million-run Monte Carlo simulation, and a convergence plot; the simulation error against the exact 0.63212 value stayed within ±0.001.
— Hy3 is free on WorkBuddy for two weeks after the official release.
Conclusions

Long-range agent tasks break most often on lazy estimation (word counts guessed instead of counted) and cross-tool data drift (PPT charts diverging from Excel source data); the self-check step in the prompt is a practical mitigation, not a formality.

High constraint density in a single prompt—colors, pixel-level shadows, hover params, corner-radius rules, inline SVGs—acts as a sharper planning test than multiple simpler prompts, because it forces the model to hold all rules simultaneously rather than process them sequentially.

Cross-referencing live web sources for model metadata is a cheap, repeatable accuracy test: the ground truth is known in advance, and the agent's willingness to report 'not found' instead of hallucinating is as informative as its extraction accuracy.

The derangement case ties reasoning to execution in a way that pure math benchmarks miss: a correct derivation followed by buggy simulation code or a mislabeled plot still fails the task, so the score reflects end-to-end reliability.

Concepts & terms
Mixture of Experts (MoE)
A model architecture where only a subset of parameters (experts) are activated per input token, reducing compute cost. Hy3 uses 192 routed experts, activating the top 8 plus one shared expert per token.
Neo-Brutalism
A web design style characterized by thick black borders, hard block shadows without blur, high-saturation color blocks, mixed corner radii, and deliberately raw, unpolished layouts.
Derangement problem
A classic combinatorics problem: given n items randomly assigned to n positions, the probability that no item ends up in its original position approaches 1/e (≈0.3679) as n grows. The complement—at least one correct assignment—approaches 1 - 1/e ≈ 0.6321.
Context compression
A technique used during long agent runs to summarize or truncate the conversation history so the model does not exceed its context window, preserving essential state while discarding redundant detail.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗