# Benchmark caveats, corrected datasets, and evaluation harnesses are documented in the source model card.
# Sources (accessed 2026-08-28):
# https://huggingface.co/Qwen/Qwen3.8-Flash-Next
# https://huggingface.co/api/models/Qwen/Qwen3.8-Flash-Next
# https://qwen.ai/blog?id=qwen3.8-flash-next
# Hub lastModified 2026-08-27T05:03:36Z is the open-weight drop (Do not use
# Hub createdAt, staged countdown page).
# Experimental preview of the Qwen4 architecture (Qwen4Exp): hybrid
# Gated DeltaNet + Qwen Sparse Attention, 512 experts (10 routed + 1 shared),
# 125B total with 6B active plus 51B n-gram embedding and 4B MTP.
# Thinking always on: reasoning_effort low|medium|xhigh (default xhigh).
# Native context 262K, extensible up to 1M tokens.

name = "Qwen3.8 Flash Next"
description = "Open-weight experimental preview of the Qwen4 architecture: hybrid-attention MoE (125B total, 6B active) with vision encoder for coding, agent tasks, and image and video understanding"
family = "qwen"
release_date = "2026-08-27"
last_updated = "2026-08-27"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = true
license = "qwen-community-1.0"

[limit]
context = 262_144
output = 131_072

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

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

[[benchmarks]]
name = "DeepSWE"
score = 58.7
metric = "score"
version = "1.1"
harness = "mini-swe-agent"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "SWE-Bench Pro"
score = 62.5
metric = "score"
harness = "Claude Code"
dataset = "Qwen refined and corrected task set"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "SWE-Bench Multilingual"
score = 81
metric = "score"
harness = "mini-swe-agent"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "NL2Repo"
score = 48.1
metric = "score"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "CoWorkBench"
score = 73.9
metric = "score"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "JobBench"
score = 55.7
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "Agents' Last Exam"
score = 24.3
metric = "pass@1"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "Agents' Last Exam"
score = 51.2
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "Toolathlon-Verified"
score = 73.5
metric = "pass@1"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "IFBench"
score = 81.3
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "GPQA Diamond"
score = 91.7
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 35.9
metric = "score"
variant = "without tools; GPT-4o judge"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "LiveCodeBench"
score = 91.9
metric = "score"
version = "6"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "ClawEval-MM"
score = 64.4
metric = "pass@3"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "ClawEval-MM"
score = 60.4
metric = "average score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "RecreationBench"
score = 49.9
metric = "score"
dataset = "internal"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "AndroidWorld"
score = 84.5
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "OSWorld"
score = 19.4
metric = "binary completion rate"
version = "2.0"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "OSWorld"
score = 52.3
metric = "partial score"
version = "2.0"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "Vision2Web"
score = 64
metric = "score"
harness = "Claude Code"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"
variant = "gpt-5.4-2026-03-05 judge"

[[benchmarks]]
name = "ERQA"
score = 72.3
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "LVBench"
score = 76.6
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "RealWorldQA"
score = 88.5
metric = "score"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "MathVision"
score = 90.6
metric = "score"
variant = "without code interpreter"
dataset = "corrected annotations"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "MathVision"
score = 95.7
metric = "score"
variant = "with code interpreter"
dataset = "corrected annotations"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 84.6
metric = "score"
variant = "without code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 90.6
metric = "score"
variant = "with code interpreter"
source = "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"