On the 2026-11-15 leaderboard, Vals AI ranked Tencent Hy4 (the production release launched on November 4) at 59.35%, placing it 17/38 overall. Its performance varied across coding, finance, financial modeling, and tax benchmarks. This is a result for a specific GDP-weighted task set, not a measure of general-purpose capability.
Tasks this result can help assess: Hy4's relative performance, cost, and runtime on finance, coding, legal, and tax work tasks in Vals Index v2.2.
Tasks this result should not be generalized to: Predicting success rates for everyday Chinese-language use, uncovered industries, specific product configurations, or all real-world workflows; it also cannot be directly compared with aggregate scores from other evaluation frameworks.
Model version covered: Tencent Hy4 production release. The Vals model page lists it as released on 2026-11-04 with open weights. These results are not for Hy4 Preview.
Test environment or client: Vals AI's Vals Index v2.2 agent evaluation. The Hy4 model page lists Tencent as the default provider; the release notes say this evaluation used OpenRouter, with costs calculated at Tencent's list prices. Different sub-benchmarks may use different providers and parameters.
Reasoning level and parameters: The model page lists temperature 0.9, top_p 1, top_k default, and a maximum output of 65,536 tokens; it does not specify Hy4's reasoning level. The page notes that some sub-benchmarks may use different parameters.
Vals Index was updated on 2026-11-15 to version v2.2. It first calculates scores for four industries separately, then weights them by their respective shares of US GDP:
Finance = average (Finance Agent v2.1, EMB v1.1)
Coding = average (Terminal-Bench 4.0, Vibe Code Bench v1.1, Code Migration v1.1)
Legal = average (Legal Research Bench, HLAB-Vals-1.1)
Tax = Tax Agent Bench
Vals Index = (8.0 × Finance + 5.6 × Coding + 1.2 × Legal + 0.5 × Tax) / 15.3The weights correspond to approximate US GDP shares of 8.0% for finance and insurance, 5.6% for information, 1.2% for legal services, and 0.5% for accounting and tax preparation services. This is an economic-impact proxy chosen by Vals, not an estimate of the share of GDP that AI could replace.
Vals says each component uses the independently published score and original scoring method of its corresponding benchmark. The exception is Code Migration v1.1: the Index uses a fixed subset of 50 CLI migration tasks and all 10 COBOL tasks, weighted 75%/25%. Vals reports a Spearman correlation of 0.99 with the full v1.1 results for this subset, a mean absolute error of 1.4 percentage points, and model rank shifts of no more than two places. Scoring methods differ across sub-benchmarks: some use automated validation or a gold standard, while others use fixed LLM judges. The Index is therefore not a single scorer that is entirely independent of judge models.
The public methodology page gives the following test sample sizes. These counts refer to benchmark tasks, not the number of models on the leaderboard. The Vals Index page does not disclose the exact test set sizes for EMB or Vibe Code Bench v1.1. The Code Migration count is the Index subset, not the full benchmark size.
| Component | Public test sample size or run setup |
|---|---|
| Finance Agent v2.1 | 449 test questions |
| EMB v1.1 | 51 test questions; each tested in Template and Scratch modes |
| Terminal-Bench 4.0 | 66 tasks; each model is run 3 times, with the mean of the three runs reported |
| Vibe Code Bench v1.1 | Exact test set size not specified on the page; evaluation is described as covering hundreds of app specifications |
| Code Migration v1.1 | Fixed Index subset: 50 CLI migrations + 10 COBOL migrations; the full version also includes 120 CLI migrations |
| Legal Research Bench | 208 test questions |
| HLAB-Vals-1.1 | 120 tasks |
| Tax Agent Bench | 193 test questions |
Overall: 59.35% (±1.08 percentage points), 17/38; one Index run cost $1.72, took 51 minutes 31 seconds, and used 17.49M tokens.
Finance: Vals Finance Index 58.05% (±1.50), 17/41. Finance Agent v2.1 scored 64.70% (±2.05), 15/41; EMB v1.1 scored 66.18% (±2.64), 33/43.
Coding: Terminal-Bench 4.0 scored 39.39% (±0.88), 19/65; Vibe Code Bench v1.1 scored 81.01% (±3.55), 34/129. The Hy4 model page does not list its Code Migration score separately.
Legal: Vals Legal Index scored 70.55% (±1.81), 18/39; Legal Research Bench scored 41.35% (±3.42), 35/95; HLAB-Vals-1.1 scored 12.50% (±2.66), 19/39.
Tax: Tax Agent Bench scored 34.40% (±3.30), 20/86.
Run specifications: The model page lists input/output prices of $0.83/$2.50 per million tokens, a 1M context window, and a 66K maximum output; leaderboard costs use Tencent's list prices. The model page reports an overall timeout fallback rate of 0% and a refusal rate of 0.14%.
Component “rank/total models” indicates a model's relative position on each benchmark leaderboard, not the number of task samples. Error terms are reported on the page as ±1 standard error; error calculation methods should not be conflated across benchmarks.
| Metric | Hy4 result | Current page basis |
|---|---|---|
| Vals Index v2.2 | 59.35% ±1.08; 17/38 | 2026-11-15 leaderboard snapshot |
| Finance Agent v2.1 | 64.70% ±2.05; 15/41 | Component on model page |
| EMB v1.1 | 66.18% ±2.64; 33/43 | Component on model page |
| Terminal-Bench 4.0 | 39.39% ±0.88; 19/65 | Component on model page |
| Vibe Code Bench v1.1 | 81.01% ±3.55; 34/129 | Component on model page |
| Legal Research Bench | 41.35% ±3.42; 35/95 | Component on model page |
| HLAB-Vals-1.1 | 12.50% ±2.66; 19/39 | Component on model page |
| Tax Agent Bench | 34.40% ±3.30; 20/86 | Component on model page |
| One Index run | $1.72; 51 minutes 31 seconds; 17.49M tokens | 2026-11-15 leaderboard snapshot; cost based on Tencent's list prices |
Vals Index's update history also retains the old leaderboard ranks from its November 4 launch. This article uses the dynamic leaderboard dated November 15 and the current Hy4 model page rankings; it does not mix in historical ranks such as 13/33 from earlier updates.
These results show Hy4 achieving a mid-range overall score on the task set selected for Vals Index, with marked variation among the coding components. The legal component also cannot be understood from the weighted legal-industry index alone. The weights for the four industries, task coverage, and scoring methods across sub-benchmarks all affect the total score; GDP weighting is simply the proxy used by Vals.
Fixed rules make it possible to compare model results within the same sub-benchmark under that benchmark's method, but this does not mean every score is produced by deterministic code: methodology pages disclose LLM judging for EMB Scratch, Legal Research, HLAB, and Tax Agent. Most test sets are also closed, and the public leaderboard does not provide Hy4's per-question outputs. Read the total and component scores together with their errors, costs, and test conditions. Neither a high score on one component nor the overall rank guarantees performance on arbitrary user tasks.
Open Vals Index in Tabbit and record Hy4's row on the 2026-11-15, v2.2 leaderboard. Open the Hy4 model page from that row and verify the release date, component scores, errors, ranks, cost, and reasoning parameters. Record the methodology, task counts, and sample splits from Vals Index and the linked official sub-benchmark descriptions. The dynamic leaderboard may change. Vals has not published all test sets or per-question run records, so this note does not claim an independent rerun or per-question recalculation.
Hy4