DataLearner logo

Hy4 previewvsHy3

Across 7 shared benchmarks, Hy4 preview leads overall: Hy4 preview wins 7, Hy3 wins 0, with 0 ties and an average score difference of +10.51.

腾讯AI实验室
Hy4 preview

腾讯AI实验室 · 2026-08-28 · Reasoning model

腾讯AI实验室
Hy3

腾讯AI实验室 · 2026-07-06 · Reasoning model

Hy4 preview7 wins(100%)(0%)0 winsHy3

Benchmark scores

Grouped by capability, sorted by largest gap within each. 7 shared benchmarks.

Coding and Software Engineer

Hy4 preview 3/3
BenchmarkHy4 previewHy3Diff
DeepSWE64.3011 / 31Thinking High (With Tools)2830 / 31Thinking High (With Tools)+36.30
SWE-Bench Pro - Public65.706 / 59Thinking High (With Tools)57.9020 / 59Thinking High (With Tools)+7.80
SWE-bench Multilingual82.903 / 27Thinking High (With Tools)75.8010 / 27Thinking High (With Tools)+7.10

AI Agent - Tool Usage

Hy4 preview 2/2
BenchmarkHy4 previewHy3Diff
Terminal-Bench 2.185.4010 / 47Thinking High (With Tools)71.7033 / 47Thinking High (With Tools)+13.70
MCP-Atlas83.705 / 40Thinking High (With Tools)79.1011 / 40Thinking High (With Tools)+4.60

General Evaluation

Hy4 preview 1/1
BenchmarkHy4 previewHy3Diff
GPQA Diamond92.3024 / 226Thinking High (No Tools)90.4038 / 226Thinking High (No Tools)+1.90

General Knowledge

Hy4 preview 1/1
BenchmarkHy4 previewHy3Diff
HLE55.4015 / 185Thinking High (With Tools)53.2021 / 185Thinking High (With Tools)+2.20

Specs

FieldHy4 previewHy3
Publisher腾讯AI实验室腾讯AI实验室
Release date2026-08-282026-07-06
Model typeReasoning modelReasoning model
ArchitectureMoEMoE
Parameters770B295B
Context length1M256K
Max outputNot availableNot available

API pricing

Prices use DataLearner records when available; missing fields are not inferred.

ItemHy4 previewHy3
Text input¥6 / 1M tokens¥1.2 / 1M tokens
Text output¥18 / 1M tokens¥4 / 1M tokens
Cache read¥0.3 / 1M tokens¥0.4 / 1M tokens

Summary

  • Hy4 previewleads in:Coding and Software Engineer (3/3), AI Agent - Tool Usage (2/2), General Evaluation (1/1), General Knowledge (1/1)

On average across the 7 shared benchmarks, Hy4 preview scores 10.51 higher.

Largest single-benchmark gap: DeepSWE — Hy4 preview 64.30 vs Hy3 28 (+36.30).

Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.