# Benchmark caveats and evaluation harnesses: https://huggingface.co/Qwen/Qwen3.8-27B#evaluation
# Sources (accessed 2026-08-15):
# https://huggingface.co/Qwen/Qwen3.8-27B
# https://huggingface.co/api/models/Qwen/Qwen3.8-27B
# https://qwen.ai/blog?id=qwen3.8
# Hub lastModified 2026-08-14T15:00:01Z is the open-weight drop.
# Do not use Hub createdAt 2026-08-05 (staged countdown page).

name = "Qwen3.8 27B"
description = "Dense 27B vision-language model for coding, agent tasks, and image and video understanding"
family = "qwen"
release_date = "2026-08-14"
last_updated = "2026-08-14"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = true

[limit]
context = 262_144
output = 32_768

[modalities]
input = ["text", "image", "video"]
output = ["text"]

[[weights]]
label = "Hugging Face"
url = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "SWE-bench Pro"
score = 61.7
metric = "resolved"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "Terminal-Bench"
score = 73
metric = "score"
version = "2.1"
harness = "Terminus"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "NL2Repo"
score = 42.3
metric = "score"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "DeepSWE"
score = 42.2
metric = "score"
version = "1.1"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "QwenSWEBench"
score = 79
metric = "avg@3"
harness = "Claude Code"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "CoWorkBench"
score = 70.7
metric = "score"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "JobBench"
score = 33.4
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "Agents' Last Exam"
score = 20.4
metric = "pass@1"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "Agents' Last Exam"
score = 42.9
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "IFBench"
score = 79.5
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "GPQA Diamond"
score = 89.2
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 30.8
metric = "score"
variant = "without tools; GPT-4o judge"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "LiveCodeBench"
score = 90.3
metric = "score"
version = "6"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "OSWorld-Verified"
score = 84.3
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "WebArena-Verified"
score = 64.8
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"
harness = "OSWorld scaffold; official WebArena-Verified grader"

[[benchmarks]]
name = "AndroidWorld"
score = 81.9
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "SWE-MM"
score = 38.6
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "RealWorldQA"
score = 85.9
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "ERQA"
score = 65.5
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "RecreationBench"
score = 47.1
metric = "score"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "ClawEval-MM"
score = 57.4
metric = "pass@3"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "ClawEval-MM"
score = 56.9
metric = "average score"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "Vision2Web"
score = 62.9
metric = "score"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"
variant = "gpt-5.4-2026-03-05 judge"

[[benchmarks]]
name = "MathVision"
score = 90
metric = "score"
variant = "without code interpreter"
dataset = "corrected annotations"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "MathVision"
score = 94.6
metric = "score"
variant = "with code interpreter"
dataset = "corrected annotations"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "BabyVision"
score = 65.7
metric = "score"
variant = "without code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "BabyVision"
score = 85.6
metric = "score"
variant = "with code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 83.7
metric = "score"
variant = "without code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"
dataset = "corrected annotations"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 90.2
metric = "score"
variant = "with code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"
dataset = "corrected annotations"

[[benchmarks]]
name = "OmniDocBench"
score = 91.1
metric = "score"
version = "1.5"
source = "https://huggingface.co/Qwen/Qwen3.8-27B"