# Hy4 preview

Tencent · Open Weight · rank 53 · bench-align-v5 · Self-hosted

> Every figure below is reproduced as its upstream source published it: nothing is modelled, estimated, interpolated or converted. `Unavailable` means no source published the value — it is never a zero. Each value carries its evidence state and the date it was observed.

Page: https://modelscale.dev/models/hy4-preview  
JSON: https://modelscale.dev/api/model/hy4-preview

## Facts

| Field | Value | Evidence |
| --- | --- | --- |
| Canonical id | `hy4-preview` | — |
| Overall score | 61.3 | Observed 2026-10-23 · source benchlm:models |
| Context window | 1M tokens | Observed 2026-10-23 · source benchlm:models |
| Release date | 2026-08-28 | Observed 2026-10-23 · source benchlm:models |
| Access type | Open Weight | — |
| Blended $/1M (75% input / 25% output) | $1.25 | Derived from the input and output rates below |

## Capability evidence

Seven axes from the ranking source. An axis the source did not score is unavailable, not zero.

| Axis | Score | Evidence |
| --- | --- | --- |
| Agentic | Unavailable | Unavailable · source benchlm:models |
| Coding | 62.6 | Observed 2026-10-23 · source benchlm:models |
| Knowledge | 81.4 | Observed 2026-10-23 · source benchlm:models |
| Reasoning | Unavailable | Unavailable · source benchlm:models |
| Multimodal & Grounded | 88.3 | Observed 2026-10-23 · source benchlm:models |
| Instruction Following | Unavailable | Unavailable · source benchlm:models |
| Math | Unavailable | Unavailable · source benchlm:models |

## Runtime service evidence

Measured values with the date they were observed. Nothing is inferred from a sibling model or a provider claim. Regional or per-endpoint measurements appear only when the API supplies them; none are modelled.

| Measurement | Value | Observed | Last good | Evidence |
| --- | --- | --- | --- | --- |
| Time to first token | Unavailable | Unavailable | Unavailable | Unavailable · source benchlm:speed |
| Throughput | 42 tok/s | 2026-10-23 | 2026-10-23 | Observed 2026-10-23 · source openrouter:endpoints |

## Endpoint and price matrix

Every published price component, including cache reads and writes.

| Component | USD | Evidence |
| --- | --- | --- |
| Input / 1M tokens | $0.83 | Observed 2026-10-23 · source openrouter:pricing |
| Output / 1M tokens | $2.50 | Observed 2026-10-23 · source openrouter:pricing |
| Cache read / 1M tokens | $0.042 | Observed 2026-10-23 · source openrouter:pricing |
| Cache write / 1M tokens | Unavailable | Unavailable · source openrouter:pricing |
| Blended / 1M (75% input / 25% output) | $1.25 | Derived — from the input and output rates above; it has no source record of its own |
| Cost per successful task (LiveBench) | Unavailable | Fetch failed · source livebench:table |

**Self-hosted listing.** Tencent publishes Hy4 preview under Apache 2.0 for self-hosting and does not publish a first-party hosted token rate for this exact checkpoint. We represent the open-weight row as self-host/free-per-token before infrastructure costs; the separate FP8 checkpoint is a quantization of the same model rather than a separately priced API SKU. The rates above are a hosted price matched from another provider, not a first-party list price.

## Workload-aware monthly cost example

10 conversations per day × 8 messages × 22 active days, 1200 input and 400 output tokens per message, no cache. Derived here from the published rates above by this site's own calculator — not a figure any source published.

| Field | Value |
| --- | --- |
| Modelled monthly cost | $3.52 |
| Modelled tokens | 2.82M |

## Benchmark record

32 matched benchmark rows with their published value, unit, and provenance.

### Knowledge

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| GPQA (Graduate-Level Google-Proof Q&A) | 92.3 | 448 questions | Graduate level | [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://arxiv.org/abs/2311.12022) |
| GPQA-D (GPQA Diamond) | 92.3 | Graduate-level science questions | Graduate level | [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://arxiv.org/abs/2311.12022) |
| HLE (Humanity's Last Exam) | 55.4 | Expert-level questions | Frontier expert level | [Humanity's Last Exam](https://lastexam.ai/) |
| HLE w/o tools (Humanity's Last Exam without tools) | 43.4 | Expert-level questions | Frontier expert level | [Introducing GPT-5.4 mini and nano](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/) |

### Coding

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Terminal-Bench 2.1 | 85.4 | Terminal-based software-agent tasks | Professional software engineering | [Terminal-Bench](https://www.tbench.ai/) |
| SWE-bench Pro | 65.7 | 1,865 repository problems | Long-horizon professional engineering | [SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?](https://arxiv.org/abs/2509.16941) |
| SWE Multilingual | 82.9 | Multilingual software-engineering tasks | Professional software engineering | [MiniMax M2.7: Early Echoes of Self-Evolution](https://www.minimax.io/news/minimax-m27-en) |
| ProgramBench (ProgramBench: Can Language Models Rebuild Programs From Scratch?) | 17.5 | 200 program reconstruction tasks | Full-repository software architecture | [ProgramBench: Can Language Models Rebuild Programs From Scratch?](https://programbench.com/static/paper.pdf) |
| PostTrain Bench | 35.6 | Post-training software-engineering tasks | Frontier software engineering | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| NL2Repo | 58.9 | Natural language to repository tasks | System-level software comprehension | [MiniMax M2.7: Early Echoes of Self-Evolution](https://www.minimax.io/news/minimax-m27-en) |
| DeepSWE | 64.3 | 113 software engineering tasks across 91 repositories and 5 languages | Long-horizon software engineering | [DeepSWE benchmark blog](https://deepswe.datacurve.ai/blog) |

### Mathematics

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Apex | 74.2 | Advanced mathematical reasoning | Frontier math reasoning | [DeepSeek-V4 Technical Report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) |

### Reasoning

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| CritPt (Critical Physics Tasks) | 16.9 | Research-level physics questions | Research-level physics reasoning | [CritPt Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/critpt) |
| STEMBench (STEMBench120: six-domain scientific research progress) | 2.59 | 120 research tasks, 20 in each of six STEM domains | Open research targets; no run has reached a full-target success | [STEMBench: six-domain model evaluation](https://www.ulam.ai/stembench) |

### Agentic

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| DRACO (Data Research and Analysis with Complex Operations) | 77.2 | Agentic data research and analysis tasks | Professional data analysis | [Claude Opus 5 System Card](https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf) |
| Terminal-Bench 2.1 | 85.4 | Terminal-based software-agent tasks | Professional software engineering | [Terminal-Bench](https://www.tbench.ai/) |
| HLE w/ tools (Humanity's Last Exam with tools) | 55.4 | Expert questions with tool use | Frontier tool-augmented reasoning | [DeepSeek-V4 Technical Report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) |
| GDPval-AA | 1678 | Agentic real-world work tasks | Professional agentic workflows | [DeepSeek-V4 Technical Report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) |
| CyberGym | 78.4 | 1,507 vulnerability analysis instances | Real-world cybersecurity | [CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale](https://www.cybergym.io/) |
| JobBench | 61.7 | 130 tasks across 35 occupations | Professional multi-source workflows | [JobBench: Aligning Agent Work With Human Will](https://arxiv.org/abs/2605.26329) |
| MCP Atlas | 83.7 | Tool-integrated agent tasks | Advanced tool use | [Introducing GPT-5.4 mini and nano](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/) |
| Toolathlon-Verified | 74.1 | Verified multi-tool workflows | Advanced tool use | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| AutomationBench | 32.1 | 600 public automation tasks | Long-horizon automation | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| Agents' Last Exam | 22.8 | Agent tasks | Advanced agentic work | [DeepSeek V4 Flash 0731 update](https://api-docs.deepseek.com/zh-cn/updates/) |
| APEX-Agents | 37.1 | Professional-services agent tasks | Long-horizon professional work | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| WideResearch | 83.9 | Open-ended research tasks | Broad research-agent workflows | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| BankerToolBench | 78.6 | Finance and banking tool-use tasks | Professional finance-agent workflows | [MiniMax M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3) |
| Terminal-Bench 2.1 (Vals) (Terminal-Bench 2.1, Vals AI run) | 55.1 | Difficult terminal tasks | Frontier agentic | [Vals AI Terminal-Bench 2.1, Vals AI run leaderboard](https://www.vals.ai/benchmarks/terminal-bench-2-1) |
| CWE-bench v1 | 53.0 | Audit and patch held-out C/C++, Go, Java, TypeScript/JavaScript, Python, and Rust repositories | Frontier agent evaluation | [CWE-bench v1 leaderboard and methodology](https://cwe-bench.com/) |
| Hermes Index (Hermes Index: agentic capability and cost per task inside Hermes Agent) | 26.1 | Four suites: Hermes Bench (150 tasks), TerminalBench 4 without its 4 GPU tasks, TerminalBench Science and SkillsBench | Frontier agentic | [Hermes Index: an agentic model leaderboard](https://portal.nousresearch.com/bench) |
| Hermes Bench (Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behind) | 58.6 | 150 tasks across 25 categories | Frontier agentic | [Hermes Bench (Hermes Index)](https://portal.nousresearch.com/bench) |

### Multimodal & Grounded

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| OfficeQA Pro | 66.2 | Document and spreadsheet tasks | Enterprise grounded reasoning | [OfficeQA Pro: An Enterprise Benchmark for End-to-End Grounded Reasoning](https://arxiv.org/abs/2603.08655) |

## Lifecycle and limitations log

Lifecycle events the source associates with this model.

No lifecycle event references this model. That is not evidence the model has no lifecycle plan — only that this source published none.

## What this profile does not claim

Values are reproduced exactly as their sources published them, in the units those sources declared; none are converted, interpolated, or averaged across providers. Any field marked unavailable was attempted and not returned. Last attempted fetch for this model's score: 2026-10-23 13:17 UTC.