# These are the Qwen3.8-Max column of the official model card, not a claim about a separately evaluated local checkpoint.
# Sources (accessed 2026-08-06):
# https://www.qwencloud.com/models/qwen3.8-max
# https://www.qianwenai.com/models/qwen3.8-max
# https://help.aliyun.com/zh/model-studio/qwen3-8-max
# https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max
# https://help.aliyun.com/zh/model-studio/pdf-understanding
# https://platform.qianwenai.com/docs/developer-guides/tool-calling/pdf-understanding
# https://docs.qwencloud.com/token-plan/personal/token-plan-personal-overview
# https://help.aliyun.com/zh/model-studio/token-plan-personal-overview
# https://help.aliyun.com/en/model-studio/token-plan-personal-overview
# https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models
# https://docs.qwencloud.com/developer-guides/text-generation/thinking
# https://docs.qwencloud.com/developer-guides/clients-and-developer-tools/opencode
# https://platform.qianwenai.com/docs/developer-guides/clients-and-developer-tools/opencode
# https://qwen.ai/blog?id=qwen3.8
# PDF input: Model Studio / 千问AI docs list only qwen3.8-max under PDF理解
# (type:file / file_url|file_data). Model pages list Image/Text/Video badges
# and separately list PDF理解 as a Completions built-in tool. Beijing-region
# availability note on help.aliyun.com; lab capability still includes pdf.

name = "Qwen3.8 Max"
description = "2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows"
family = "qwen"
release_date = "2026-08-03"
last_updated = "2026-08-03"
attachment = true
reasoning = true
temperature = true
tool_call = true
open_weights = false

[limit]
context = 1_000_000
output = 131_072

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

[[benchmarks]]
name = "Terminal-Bench"
score = 86.6
metric = "avg@10"
version = "2.1"
harness = "Claude Code"
variant = "5h timeout"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "SWE-Bench Pro"
score = 67.7
metric = "resolved"
harness = "Claude Code"
dataset = "Qwen refined and corrected task set"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "DeepSWE"
score = 56.6
metric = "score"
version = "1.1"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "NL2Repo"
score = 55.9
metric = "score"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "FrontierSWE"
score = 73.5
metric = "dominance score"
date = "2026-08-03"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "MLS-Bench-Lite"
score = 41
metric = "score"
harness = "Claude Code"
variant = "5h timeout"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "PaperBench"
score = 93
metric = "score"
harness = "BasicAgent"
variant = "Code-Dev; 3 runs; 12h timeout; Opus 4.6 judge"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "AndroidBench"
score = 75.1
metric = "avg@3"
dataset = "95 public tasks"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "QwenSWEBench"
score = 80.7
metric = "avg@3"
harness = "Claude Code"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "QwenQoderBench"
score = 58.4
metric = "avg@5"
harness = "Claude Code"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "QwenReactBench"
score = 1724
metric = "Elo"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "QwenSVGBench"
score = 1713
metric = "Elo"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "CoWorkBench"
score = 74.8
metric = "score"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "WorkSpaceBench"
score = 67.7
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "JobBench"
score = 53.4
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "SkillsBench"
score = 70.2
metric = "avg@3"
version = "1.1"
dataset = "87 tasks"
harness = "OpenCode"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "Agents' Last Exam"
score = 27
metric = "pass@1"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "Agents' Last Exam"
score = 52.4
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "AutomationBench"
score = 27.3
metric = "pass@1"
dataset = "600 public tasks"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "Toolathlon-Verified"
score = 72.5
metric = "pass@1"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "WideSearch"
score = 81.9
metric = "average item F1"
harness = "Qwen-Agent"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 56.2
metric = "score"
variant = "with tools"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 43.6
metric = "score"
variant = "without tools"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "GPQA Diamond"
score = 92.6
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "IFBench"
score = 82.8
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "$OneMillion-Bench"
score = 52.5
metric = "expert score"
variant = "gemini-3.1-pro-preview judge"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "HealthBench"
score = 60.2
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "PLawBench"
score = 73.2
metric = "score"
variant = "gemini-3.1-pro-preview judge"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "PRBench"
score = 57.6
metric = "score"
variant = "legal"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "PRBench"
score = 58.3
metric = "score"
variant = "finance"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "MRCR"
score = 92.9
metric = "score"
version = "2"
dataset = "256K; 8-needle"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

[[benchmarks]]
name = "LongBench"
score = 66.3
metric = "score"
version = "2"
source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"