← 模型目录

Hy3

Tencent / 腾讯混元 · 2026-07-06 · 类别未确认

官方发布来源 · 按需求选型 · 查看覆盖缺口

Hy3

腾讯发布混元 Hy3：295B 总参/21B 激活 MoE、256K 上下文、三种思考模式，Apache 2.0 开源。评测覆盖编码、智能体搜索与推理，亮点如 SWE-bench Verified 78%、BrowseComp 84.2%。

输入模态: 文本
上下文: 256K
参数: 295B-A21B MoE
价格（每百万 tokens）: CNY 输入 1 / 输出 4 · 页面原文为 1/4/0.25 元/百万 tokens 三档报价，第三档含义未标注

本变体的评测证据

mrcr 75.1 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 更可靠的产品体验 → 复杂上下文承接与多轮意图保持能力 · row: MRCR · quote_snippet: 长对话理解基准中取得显著跨越（如 MRCR 从 42.9% 升至 75.1%）

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Page prose value (DOM text). The 42.9% in the same sentence refers to Hy3 preview, not a competitor. Long-context understanding benchmark family; page does not print the MRCR variant/needle count — variant unspecified.

打开官方来源

swebench 官方未公布数值 模型 hy3 · 版本 Verified · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 更可靠的产品体验 → 输出格式和工具调用稳定性 · row: SWE Bench Verified · quote_snippet: 在 SWE Bench Verified 上的分数标准差控制在 4 个百分点以内

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Prose mentions SWE Bench Verified by name but reports no score for Hy3; the disclosure is a cross-scaffold robustness claim (std < 4pp across Codebuddy / Cline / KiloCode scaffolds). Score value appears only in the appendix image (see the separate pending image row).

打开官方来源

swebench-multilingual 75.8 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: swebench-multilingual · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "SWE-agent scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

swebench 78 模型 hy3 · 版本 Verified · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: swebench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "SWE-agent scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

swebench-pro 57.9 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: swebench-pro · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "SWE-agent scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

terminalbench 71.7 模型 hy3 · 版本 2.1 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: terminalbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "Terminus-2 scaffold, xml parser, 4h agent timeout, 16 cores / 32 GB, max 500 episodes",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

nl2repo 45.6 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: nl2repo · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "Claude Code, 250-turn budget, 12000s timeout, 4 CPU / 32 GB, anti-hacking constraints",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

deepswe 28 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: deepswe · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "mini-swe-agent, 2h timeout per task, 2 CPU / 8 GB, strict network isolation",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-backend-2-0 25 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-backend-2-0 · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "Claude Code scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-backend-2-0 not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-swe-max 49 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-swe-max · figure: data/model-releases/official/_archive/tencent/hy3/images/split/01.png

{
  "harness": "Claude Code scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-swe-max not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/01.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-companybench 41.7 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-companybench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "Claude Code scaffold",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-companybench not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

browsecomp 84.2 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: browsecomp · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "internal harness; self-summary context management (agentic search)",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

widesearch 76.4 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: widesearch · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "internal harness (agentic search)",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

deepsearchqa 91 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: deepsearchqa · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "internal harness (agentic search)",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

mcp-atlas 79.1 模型 hy3 · 版本 500-task public set · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: mcp-atlas · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "Scale April 2026 methodology, 100 tool-call budget per task",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": "Gemini 2.5 Pro"
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

toolathlon 48.5 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: toolathlon · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

apex-agents 25.6 模型 hy3 · 版本 pass@1 · 指标 pass_at_1 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: apex-agents · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

claw-eval 68.5 模型 hy3 · 版本 pass^3 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: claw-eval · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "internal harness, 20260325 version (105 queries)",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": "Gemini-3.5-flash"
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

wildclawbench 53.6 模型 hy3 · 版本 text-only (35 queries) · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: wildclawbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "OpenClaw Harness",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

skillsbench 55.3 模型 hy3 · 版本 79-task text-only subset, mean of 3 runs · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: skillsbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/02.png

{
  "harness": "Claude Code",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": 3,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/02.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

e-bench 50.2 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: e-bench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: e-bench not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-finmodelbench 69 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-finmodelbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-finmodelbench not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

prodbench 23 模型 hy3 · 版本 pass^3 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: prodbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": "OpenClaw Harness",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: prodbench not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-skillsworld 45.8 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-skillsworld · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-skillsworld not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hlehle 53.2 模型 hy3 · 版本 with tools, text-only · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hlehle · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-euler-pro 24.2 模型 hy3 · 版本 with tools · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-euler-pro · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-euler-pro not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

gpqa 90.4 模型 hy3 · 版本 Diamond · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: gpqa · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hlehle 37 模型 hy3 · 版本 no tools, text-only · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hlehle · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

frontier-science-research 21.3 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: frontier-science-research · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

frontier-science-olympiad 74.8 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: frontier-science-olympiad · figure: data/model-releases/official/_archive/tencent/hy3/images/split/03.png

{
  "harness": "self-evaluated with OpenAI FrontierScience paper judge prompts",
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": "gpt-oss-120b (high reasoning effort)"
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/03.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

usamo-2026 72 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: usamo-2026 · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

apex 38.7 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: matharena-apex · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: matharena-apex not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

arxivmath 52.2 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: arxivmath · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: arxivmath not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

horizonmath 7.1 模型 hy3 · 版本 pass@12 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: horizonmath · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

hy-math 60.9 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: hy-math · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: hy-math not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

phybench 77.4 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: phybench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

cmt-benchmark 37.8 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: cmt-benchmark · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: cmt-benchmark not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

imo-answerbench 90 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: imo-answerbench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

superchem 54.9 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: superchem · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: superchem not yet in data/benchmarks/. Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

cl-bench 23.8 模型 hy3 · 版本未说明 · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: cl-bench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/04.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/04.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

cl-bench 17 模型 hy3 · 版本 life · 指标 resolved_rate · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: cl-bench · figure: data/model-releases/official/_archive/tencent/hy3/images/split/05.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/05.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源

aa-lcr 73.4 模型 hy3 · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 附录：模型得分 · row: aa-lcr · figure: data/model-releases/official/_archive/tencent/hy3/images/split/05.png

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "highest tier (per appendix note: reasoning effort set to the highest tier for all models)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Value read from the page appendix image and confirmed by re-reading the archived image with the Read tool on 2026-09-01 (data/model-releases/official/_archive/tencent/hy3/images/split/05.png). Page prints the row under the Hy3 column; protocol fields transcribed from the notes block printed beneath the same table.

打开官方来源