Alibaba Cloud · Qwen3.8 · Reviewed 2026-09-04

Qwen3.8-Flash

Qwen's production multimodal Flash API model, based on the Qwen3.8-Flash-Next open-weight architecture preview.

All models

Overall · 0–100—

Status: active
Released: 2026-08-26
Context: 1,000,000 tokens
Parameters: 125B total · 6B active
Inputs: Text · Image · Video
Type: multimodal
Max output: 131,072 tokens
API model ID: qwen3.8-flash
Reasoning: low / medium / high / xhigh
Knowledge cutoff: —
Training cutoff: —
Weights / license: Unavailable
Price / 1M tokens: $0.113 in · $0.382 out
Cached input: $0.014

Independent benchmark scores

No snapshot score

Coding

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Reasoning

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Research

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Multimodal

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Agentic

—

No comparable score for this task lens in the selected independent snapshot.

Raw benchmark results

Source-backed rows with the original evaluation context preserved.

24 results

Benchmark	Score	Source	Evaluation notes
Agents' Last ExamEvaluated 2026-08-26	51.2%	Qwen / Hugging Face	Qwen model-card result; overall score=51.2 and Pass@1=24.3; official protocol.
AndroidWorldEvaluated 2026-08-26	84.5%	Qwen / Hugging Face	Qwen model-card AndroidWorld result.
Artificial Analysis Intelligence IndexEvaluated 2026-09-03	56	Artificial Analysis	Artificial Analysis Intelligence Index v4.1.1; score is for the related Qwen3.8-Flash-Next open-weight artifact, not a claim that the production API and checkpoint are identical.
CharXiv RQEvaluated 2026-08-26	90.6%	Qwen / Hugging Face	Qwen model-card CharXiv RQ result with CI; without CI=84.6.
ClawEval-MMEvaluated 2026-08-26	60.4%	Qwen / Hugging Face	Qwen model-card ClawEval-MM average score; Pass@3=64.4.
CoWorkBenchEvaluated 2026-08-26	73.9%	Qwen / Hugging Face	Qwen in-house CoWorkBench result for long-horizon office and productivity agent tasks.
DeepSWE v1.1Evaluated 2026-08-26	58.7%	Qwen / Hugging Face	Qwen model-card result; DeepSWE v1.1, highest score across Claude Code and mini-SWE-agent harnesses.
ERQAEvaluated 2026-08-26	72.3%	Qwen / Hugging Face	Qwen model-card ERQA result.
GPQA / GPQA DiamondEvaluated 2026-08-26	91.7%	Qwen / Hugging Face	Qwen model-card GPQA Diamond result.
Humanity's Last Exam (HLE)Evaluated 2026-08-26	35.9%	Qwen / Hugging Face	Qwen model-card HLE result; judged by GPT-4o.
IFBenchEvaluated 2026-08-26	81.3%	Qwen / Hugging Face	Qwen model-card IFBench result.
JobBenchEvaluated 2026-08-26	55.7%	Qwen / Hugging Face	Qwen model-card JobBench result.
LiveCodeBench v6Evaluated 2026-08-26	91.9%	Qwen / Hugging Face	Qwen model-card LiveCodeBench v6 result.
LVBenchEvaluated 2026-08-26	76.6%	Qwen / Hugging Face	Qwen model-card LVBench result.
MathVisionEvaluated 2026-08-26	95.7%	Qwen / Hugging Face	Qwen model-card MathVision result with CI; without CI=90.6.
NL2RepoEvaluated 2026-08-26	48.1%	Qwen / Hugging Face	Qwen model-card result; Claude Code harness with repository-access reward-hacking controls.
OSWorld 2.0 BinaryEvaluated 2026-08-26	19.4%	Qwen / Hugging Face	Qwen model-card OSWorld 2.0 binary score; partial score=52.3.
OSWorld 2.0 PartialEvaluated 2026-08-26	52.3%	Qwen / Hugging Face	Qwen model-card OSWorld 2.0 partial score; binary score=19.4.
RealWorldQAEvaluated 2026-08-26	88.5%	Qwen / Hugging Face	Qwen model-card RealWorldQA result.
RecreationBenchEvaluated 2026-08-26	49.9%	Qwen / Hugging Face	Qwen in-house RecreationBench result.
SWE-bench MultilingualEvaluated 2026-08-26	81%	Qwen / Hugging Face	Qwen model-card result; mini-SWE-agent harness.
SWE-bench ProEvaluated 2026-08-26	62.5%	Qwen / Hugging Face	Qwen model-card result; Claude Code harness, 256K context, refined benchmark tasks.
Toolathlon VerifiedEvaluated 2026-08-26	73.5%	Qwen / Hugging Face	Qwen model-card Toolathlon Verified result; Pass@1.
Vision2WebEvaluated 2026-08-26	64%	Qwen / Hugging Face	Qwen model-card Vision2Web average result; Claude Code harness and GPT-5.4 judge.

Strengths

1M-token production context
125B total and 6B activated language-model parameters
Strong coding, agentic, and multimodal benchmark coverage

Limitations

The production service and open-weight Flash-Next artifact are related but distinct records
Parameter accounting includes separately reported n-gram and MTP components

Model evidence · official · Checked 2026-09-04

Qwen3.8-Flash model information

Official Alibaba Cloud model page for Qwen3.8-Flash production API context, modalities, 1M context limit, output limits, capabilities, pricing, and rate limits.

Model evidence · official · Checked 2026-09-04

Qwen3.8-Flash-Next model card

Official Qwen open-weight model card containing architecture, parameter breakdown, context extension, sampling guidance, and benchmark tables. Qwen documents Qwen3.8-Flash as the production version based on this preview.

Model evidence · official · Checked 2026-09-04

Qwen3.8-Flash-Next: A new architecture, towards ultimate cost-efficiency

Official Qwen release context distinguishing the open-weight Flash-Next architecture preview from the production Qwen3.8-Flash service.

Capability Index source

Artificial Analysis Intelligence Index v4.3.2

Underlying raw metrics and exact model configurations remain authoritative.

Capability Index source

Artificial Analysis methodology

Underlying raw metrics and exact model configurations remain authoritative.

Task lenses preserve raw scores. Radar axes use score / current leader × 100; overall is the equal-weight average of all five indices. Missing evidence is not a zero or a low rank.