Hy4 Benchmark — Scores & Architecture

Benchmark results for Tencent Hy4 Preview across coding, search, working-agent, STEM, and reasoning tasks, plus an architecture-level comparison. Last updated: August 28, 2026.

How to Read This Dataset

The tables below are transcribed from benchmark-data.md, a dataset compiled from public reports, provider disclosures, and clearly labeled third-party or internal evaluations. They contain 46 benchmark rows across five task families and eight model columns. The Hy4 Preview column is highlighted in blue. An asterisk comes from the source party's own testing, a slash preserves multiple reported values, and a dash means no value was shown.

Important: These results are provided for reference and are not universally verified official scores. Results use different harnesses, reasoning settings, judges, resource limits, and run counts. Compare models within the same benchmark and read the evaluation notes before drawing conclusions across task families.

Agentic Coding

Scores transcribed from benchmark-data.md; source markers are preserved.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
SWE-bench Multilingual75.882.977.3*82.6*81.3*80.8*74.1*89.5/85.8*
SWE-bench Pro57.965.760.3*67.7/61.6*64.6*63.3*64.6/60.5*79.2/79.9*
DeepSWE28.064.362.7/58.8*56.6/55.6*66.9/68.1*67.5/74.0*72.7/68.9*68.8/74.7*
SWE Atlas - Codebase Q&A30.864.053.4*55.4*55.8*35.2*58.1*57.5*
SWE Atlas - Test Writing35.957.845.6*52.8*49.6*35.6*49.6*68.7*
SWE Atlas - Refactoring32.953.348.6*51.0*51.9*37.4*52.4*60.0*
SWE-Marathon5.031.919.0*31.0*42.5/35.6*42/44.4*42.5/35.9*50.0/48.0*
Terminal-Bench 2.170.885.487.9/80.3*86.6/85.8*88.2/88.3*88.3/85.7*88.8/88.3*86.7/85.4*
NL2Repo-Bench45.658.961.5/54.4*55.9/58.0*58.0/56.1*58.0/58.3*56.8*75.3*
CyberGym51.878.483.3/80.3*78.5/78.5*84.5/83.0*80.083.6-
ProgramBench3.017.515.5*17.5*18.0*24.5*25.0*39.5*
PostTrainBench V1.114.535.624.5*-33.2*32.0*36.235.0
Harbor-Index15.639.636.9*38.8*42.5*-46.3*56.9*
Hy-Backend 2.0 (Internal)26.235.233.8*34.9*41.9*37.9*49.6*40.3*
Hy-SWE Max Verified (Internal)49.064.265.7*65.2*67.2*65.6*69.8*70.1*
Hy-CompanyBench V2 (Internal)29.862.464.4*63.3*64.5*63.3*70.6*72.7*

Agentic Search

Scores transcribed from benchmark-data.md; source markers are preserved.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
WideSearch81.983.981.8*81.9/81.1*83.2*81.0*86.3*84.0*
OneMillionBench (with tools)51.565.462.0*63.1*64.5*63.5*67.1*68.1*
DRACO65.277.277.3*76.4*78.1*77.5*77.7*88.6/79.2*
Hy-Live-Search (Internal)38.949.246.9*47.5*49.2*45.8*63.4*56.1*
Hy-BrowseComp-Pro2 (Internal)55.056.146.5*46.7*48.4*58.1*56.4*61.3*

Working Agent

Scores transcribed from benchmark-data.md; source markers are preserved.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
OfficeQA Pro54.166.265.4*65.4*66.2*65.4*65.4*66.9/66.9*
MCP-Atlas (public)75.083.782.5*81.9*81.9*84.2/82.8*82.5*85.7*
Toolathlon-Verified56.274.174.1/70.1*72.5/69.1*73.0/73.8*76.5/74.7*73.2*76.5*
APEX-Agents (pass@1)24.437.132.4*34.0*38.1*41.0/37.2*39.9/37.9*41.8*
SkillsBench (79, text-only)55.362.965.0*66.7*63.3*51.9*62.5*63.7*
JobBench34.661.754.1*53.0/52.2*58.2*54.3/54.7*45.4/46.8*68.0*
WorkspaceBench58.260.265.4*67.7/66.9*68.2*65.0*65.3*75.0*
Agents' Last Exam (ALE-CLI)17.122.821.9*25.4*23.8*23.2*27.6*25.1*
GDPval-AA V2 (Elo, official)12131678158017171763167517111831
AutomationBench (v1.0.6)16.132.130.4*39.8/41.1*48.2/49.4*46.7/45.5*45.8/39.9*48.7*
BankerToolBench68.878.673.1*74.7*77.8*73.5*79.0*81.9*
E-Bench (Internal)48.577.161.3*66.8*71.4*73.8*80.6*77.8*
E-Bench-Code (Internal)64.479.064.3*67.1*66.5*77.6*83.3*82.7*
Hy-FinAgentBench (Internal)69.579.778.5*77.2*80.4*78.5*83.0*82.0*
Hy-FinmodelBench v2 (Internal)28.657.051.3*52.5*57.8*52.4*65.3*66.0*

STEM Agent

Scores transcribed from benchmark-data.md; source markers are preserved.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
BioMysteryBench54.971.361.6*58.9*69.0*61.3*73.1*72.1*
HLE (with tools, text-only)51.955.460.0/55.8*56.2/54.1*62.5/54.3*57.0*60.2*60.9*

Reasoning

Scores transcribed from benchmark-data.md; source markers are preserved.

BenchmarkHy3Hy4 previewDeepSeek V4 Pro 0813Qwen 3.8 MaxGLM 5.3Kimi K3GPT 5.6 SolClaude Opus 5
CritPt (official)4.916.918.020.019.123.432.329.1
GPQA Diamond90.992.392.8/91.7*92.6/92.2*91.7/91.4*93.5/92.8*94.1/94.7*93.7/93.3*
HLE (no tools, text-only)34.443.442.7/40.5*43.6/41.5*42.346.9/46.6*49.5/49.6*54.9/53.2*
SUPERChem52.666.462.0*61.9*58.5*66.9*73.6*76.7*
ArXivMath51.766.662.1*67.1*-60.8*79.5*71.5*
HorizonMath (pass@4)3.58.84.42*5.31*-7.08*10.62*5.3*
MathArena Apex 202538.774.266.3*72.8*-68.4*90.0*91.4*
BrokenArXiv26.754.643.1*42.7*-56.3*64.4*77.7*

Architecture-Level Comparison

Both models are described as activating 49B parameters per token, but their published designs use different depths and routing choices.

Architecture detailHy4 PreviewDeepSeek-V4-Pro
Total / Active parameters770B / 49B1.6T / 49B
Activation ratio≈6.4%≈3.1%
Backbone layers7861
Layer structure1 dense FFN + 77 MoE0–1 HCA + 2–60 alternating CSA/HCA
Routed experts per MoE layer256384
Shared experts per layer11
Experts activated per tokentop-8 + sharedtop-6 + shared
Speculative decoding1 native MTP layer (10B / 0.7B active)1 MTP layer; DSpark module (GA only)
Context window1M1M
Max output64K384K
Open weightsNo (hosted API only)Yes (MIT)
What this means: Hy4 spreads active parameters over fewer, larger experts, while DeepSeek V4 uses more routed experts and finer routing. Both designs should be evaluated against task-level results.

What Each Benchmark Family Measures

FamilyTypical focusExamples in this dataset
Agentic CodingIssue resolution, code generation, testing, terminal workSWE-bench, DeepSWE, Terminal-Bench, CyberGym
Agentic SearchResearch, retrieval, browsing, and tool-assisted discoveryWideSearch, DRACO, Hy-Live-Search
Working AgentOffice tasks, tools, automation, and applied workflowsMCP-Atlas, Toolathlon, GDPval-AA, E-Bench
STEM AgentScientific investigation and tool-supported knowledge workBioMysteryBench, HLE with tools
ReasoningExpert knowledge, mathematics, and abstract problem solvingGPQA, SUPERChem, ArXivMath, MathArena

Evaluation Notes

  • Each model was evaluated and reported at the highest available reasoning setting; results marked with `*` are from the source party's own testing.
  • Hy3 benchmark scores may differ from previously reported results because of updates to the evaluation harness, judge model, and anti-hacking mechanisms.
  • SWE-bench Multilingual and SWE-bench Pro use different evaluation methods.
  • For DeepSWE, the mini-swe-agent was used; each task ran in a sandbox with 8 CPU cores and 16 GB memory.
  • SWE-Marathon used the official evaluation setup, with tasks rerun until a valid score was obtained.
  • Terminal-Bench 2.1 used the Claude Code harness, with up to 500 turns and a 12-hour timeout per trial; resources were capped at 16 CPU cores and 32 GB RAM.
  • MCP-Atlas followed the April 2026 methodology, using a 500-task public set, a 100 tool-call budget per task, and Gemini 3.1 Pro Preview as the judge model.
  • SkillsBench results are averaged across three runs.
  • SWE-Atlas used the Claude Code harness with a 200-turn budget per task.
  • NL2Repo-Bench used a 100-turn budget with anti-hacking constraints and tool-call monitoring.
  • ProgramBench used a 2,000-turn budget; the DeepSeek-V4-Pro result may have been affected by repeated loops.
  • Hy-SWE Max Verified (Internal) used an internally built 300-task benchmark with WebSearch and WebFetch disabled.
  • Agents' Last Exam used 105 ALE-CLI tasks and the official evaluation protocol.
  • Apex-Agents used a 250-step trial with resources capped at 4 CPU cores and 16 GB RAM.
  • BioMysteryBench used an agent harness with a two-hour and 250-turn budget per task.
  • GPQA used each model's highest reasoning setting.
  • All results should be interpreted together with each benchmark's definition, evaluation environment, and number of runs; they should not be treated as an absolute ranking across benchmarks.

Frequently Asked Questions

What are the Hy4 benchmark scores?

The supplied dataset reports Hy4 Preview scores across Agentic Coding, Agentic Search, Working Agent, STEM Agent, and Reasoning tasks. Values are shown in the tables, including 82.9 on SWE-bench Multilingual, 85.4 on Terminal-Bench 2.1, 83.7 on MCP-Atlas, 71.3 on BioMysteryBench, and 92.3 on GPQA Diamond.

How does Hy4 compare to other models on benchmarks?

Hy4 Preview leads the supplied model set on several listed tasks, while other models lead on others. Because harnesses and settings differ, compare models within the same benchmark rather than treating the dataset as one absolute ranking.

How does Hy4's architecture compare to DeepSeek V4?

Hy4 Preview uses 78 layers with 256 routed experts per layer and top-8 routing (770B total / 49B active). DeepSeek-V4-Pro uses 61 layers with 384 routed experts per layer and top-6 routing (1.6T total / 49B active). Both activate 49B parameters per token.

When will Hy4 be officially released?

Hy4 is currently in grey-testing and Tencent has not announced a specific release date.

Which benchmarks matter most for Hy4?

The most useful rows depend on your workload: SWE-bench and Terminal-Bench for coding, MCP-Atlas and Toolathlon for tool use, BioMysteryBench and HLE for STEM and research, and GPQA Diamond or MathArena for reasoning.

Test Hy4 Preview Yourself →

Read the Hy4 Preview home page or the Hy4 vs DeepSeek comparison.