Hy4 previewvsDeepSeek-V4-Pro
Across 11 shared benchmarks, Hy4 preview leads overall: Hy4 preview wins 6, DeepSeek-V4-Pro wins 4, with 1 ties and an average score difference of +7.53.
Hy4 preview
腾讯AI实验室 · 2026-08-28 · Reasoning model
DeepSeek-V4-Pro
DeepSeek-AI · 2026-08-13 · Reasoning model
Hy4 preview6 wins(55%)Ties1(36%)4 winsDeepSeek-V4-Pro
Benchmark scores
Grouped by capability, sorted by largest gap within each. 11 shared benchmarks.
AI Agent - Tool Usage
DeepSeek-V4-Pro 2/4| Benchmark | Hy4 preview | DeepSeek-V4-Pro | Diff |
|---|---|---|---|
| CyberGym | 78.403 / 4Thinking High (With Tools) | 83.302 / 4极高强度思考（工具） | -4.90 |
| Terminal-Bench 2.1 | 85.4010 / 47Thinking High (With Tools) | 87.906 / 47极高强度思考（工具） | -2.50 |
| AutomationBench | 32.103 / 10Thinking High (With Tools) | 31.804 / 10极高强度思考（工具） | +0.30 |
| Toolathlon-Verified | 74.103 / 8Thinking High (With Tools) | 74.103 / 8极高强度思考（工具） | — |
Coding and Software Engineer
Hy4 preview 3/4| Benchmark | Hy4 preview | DeepSeek-V4-Pro | Diff |
|---|---|---|---|
| SWE-Bench Pro - Public | 65.706 / 59Thinking High (With Tools) | 52.1042 / 59Normal (With Tools) | +13.60 |
| SWE-bench Multilingual | 82.903 / 27Thinking High (With Tools) | 69.8021 / 27Normal (With Tools) | +13.10 |
| NL2Repo-Bench | 58.902 / 11Thinking High (With Tools) | 61.501 / 11极高强度思考（工具） | -2.60 |
| DeepSWE | 64.3011 / 31Thinking High (With Tools) | 62.7013 / 31极高强度思考（工具） | +1.60 |
Agent Level Benchmark
DeepSeek-V4-Pro 1/1| Benchmark | Hy4 preview | DeepSeek-V4-Pro | Diff |
|---|---|---|---|
| Agents' Last Exam | 22.8013 / 14Thinking High (With Tools) | 25.7010 / 14极高强度思考（工具） | -2.90 |
General Evaluation
Hy4 preview 1/1| Benchmark | Hy4 preview | DeepSeek-V4-Pro | Diff |
|---|---|---|---|
| GPQA Diamond | 92.3024 / 226Thinking High (No Tools) | 72.90147 / 226Normal (No Tools) | +19.40 |
General Knowledge
Hy4 preview 1/1| Benchmark | Hy4 preview | DeepSeek-V4-Pro | Diff |
|---|---|---|---|
| HLE | 55.4015 / 185Thinking High (With Tools) | 7.70169 / 185Normal (No Tools) | +47.70 |
Specs
| Field | Hy4 preview | DeepSeek-V4-Pro |
|---|---|---|
| Publisher | 腾讯AI实验室 | DeepSeek-AI |
| Release date | 2026-08-28 | 2026-08-13 |
| Model type | Reasoning model | Reasoning model |
| Architecture | MoE | MoE |
| Parameters | 770B | 1.6T |
| Context length | 1M | 1M |
| Max output | Not available | 384K |
API pricing
Prices use DataLearner records when available; missing fields are not inferred.
| Item | Hy4 preview | DeepSeek-V4-Pro |
|---|---|---|
| Text input | ¥6 / 1M tokens | $0.435 / 1M tokens |
| Text output | ¥18 / 1M tokens | $0.87 / 1M tokens |
| Cache read | ¥0.3 / 1M tokens | $0.003625 / 1M tokens |
Summary
- Hy4 previewleads in:Coding and Software Engineer (3/4), General Evaluation (1/1), General Knowledge (1/1)
- DeepSeek-V4-Proleads in:AI Agent - Tool Usage (2/4), Agent Level Benchmark (1/1)
On average across the 11 shared benchmarks, Hy4 preview scores 7.53 higher.
Largest single-benchmark gap: HLE — Hy4 preview 55.40 vs DeepSeek-V4-Pro 7.70 (+47.70).
Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.