# Scores and detailed evaluation recipes: https://huggingface.co/tencent/Hy3/blob/main/assets/benchmark-appendix.png
# https://cloud.tencent.com/document/product/1823/130051
name = "Hy3"
description = "Tencent Hy reasoning model for coding, instruction following, and agent tasks"
family = "Hy"
release_date = "2026-07-06"
last_updated = "2026-07-06"
attachment = false
reasoning = true
temperature = true
tool_call = true
open_weights = true

[limit]
context = 256_000
input = 192_000
output = 128_000

[modalities]
input = ["text"]
output = ["text"]

[[weights]]
label = "Hugging Face"
url = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "SWE-Bench Verified"
score = 78
metric = "resolved"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "SWE-Bench Multilingual"
score = 75.8
metric = "score"
variant = "highest reasoning effort"
harness = "SWE-agent"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "SWE-Bench Pro"
score = 57.9
metric = "score"
variant = "highest reasoning effort"
harness = "SWE-agent"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "Terminal-Bench"
score = 71.7
metric = "score"
variant = "highest reasoning effort; 4h timeout; 500 episodes"
version = "2.1"
harness = "Terminus 2"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "NL2Repo"
score = 45.6
metric = "score"
variant = "highest reasoning effort; 250 turns; 12000s timeout"
harness = "Claude Code"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "DeepSWE"
score = 28
metric = "score"
variant = "highest reasoning effort; 2h timeout"
harness = "mini-swe-agent"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "BrowseComp"
score = 84.2
metric = "score"
variant = "highest reasoning effort"
harness = "Tencent internal search harness"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "WideSearch"
score = 76.4
metric = "score"
variant = "highest reasoning effort"
harness = "Tencent internal search harness"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "DeepSearchQA"
score = 91
metric = "score"
variant = "highest reasoning effort"
harness = "Tencent internal search harness"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "MCP Atlas"
score = 79.1
metric = "score"
variant = "highest reasoning effort; April 2026; 100 tool calls"
dataset = "500 public tasks"
harness = "Scale AI"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "Toolathlon"
score = 48.5
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "APEX-Agents"
score = 25.6
metric = "pass@1"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "ClawEval"
score = 68.5
metric = "pass@3"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"
harness = "Tencent internal harness"
version = "20260325"
dataset = "105 queries"

[[benchmarks]]
name = "WildClawBench"
score = 53.6
metric = "score"
variant = "highest reasoning effort"
dataset = "35 text-only tasks"
harness = "OpenClaw"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "SkillsBench"
score = 55.3
metric = "average over 3 runs"
variant = "highest reasoning effort"
dataset = "79 text-only tasks"
harness = "Claude Code"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 53.2
metric = "score"
variant = "highest reasoning effort; with tools"
dataset = "text-only"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 37
metric = "score"
variant = "highest reasoning effort; without tools"
dataset = "text-only"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "GPQA Diamond"
score = 90.4
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "FrontierScience"
score = 21.3
metric = "score"
variant = "highest reasoning effort; research"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "FrontierScience"
score = 74.8
metric = "score"
variant = "highest reasoning effort; olympiad"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "USAMO"
score = 72
metric = "score"
variant = "highest reasoning effort"
version = "2026"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "MathArena Apex"
score = 38.7
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "ArxivMath"
score = 52.2
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "HorizonMath"
score = 7.1
metric = "pass@12"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "PHYBench"
score = 77.4
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "CMT Benchmark"
score = 37.8
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "IMOAnswerBench"
score = 90
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "SuperChem"
score = 54.9
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "CL-bench"
score = 23.8
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "CL-bench-life"
score = 17
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"

[[benchmarks]]
name = "AA-LCR"
score = 73.4
metric = "score"
variant = "highest reasoning effort"
source = "https://huggingface.co/tencent/Hy3"