Qwen3.8 Flash Next

Name: Qwen3.8 Flash Next
Author: Alibaba

Multimodal

Alibaba

Qwen3.8 Flash Next is an open-weight multimodal MoE from Alibaba's Qwen team, released August 2026 as an early preview of the Qwen4 architecture. A 125B-parameter backbone with 51B N-gram embeddings and a 4B multi-token prediction module activates only 6B parameters per token. It introduces a GDN + Qwen Sparse Attention hybrid, Gated Residual and N-gram Embedding, and reportedly costs about one-ninth of Qwen3.7-Plus to train. Native 262K context, extensible to 1M with YaRN. Licensed under the Qwen community license (not Apache-2.0).

Key Specifications

Parameters

180.0B

Context

1.0M

Release Date

August 26, 2026

Average Score

67.1%

API Documentation Research Paper Repository Model Weights Results Blog

Timeline

Key dates in the model's history

Announcement

August 26, 2026

Last Update

August 27, 2026

Today

September 8, 2026

Technical Specifications

Parameters

180.0B

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)

$0.16

Output (per 1M tokens)

$0.47

Max Input Tokens

1.0M

Max Output Tokens

65.5K

Supported Features

Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis

GPQA Diamond

GPQA Diamond, scientific reasoning • Self-reported

91.7%

Other Tests

Specialized benchmarks

DeepSWE 1.1

Agentic coding, Claude Code and mini-SWE-agent harnesses, temp=1.0, top_p=0.95, 256K context • Self-reported

58.7%

SWE-bench Pro

Claude Code harness, temp=1.0, top_p=0.95, 256K context • Self-reported

62.5%

SWE-bench Multilingual

mini-SWE-agent harness, temp=1.0, top_p=0.95, 256K context • Self-reported

81.0%

LiveCodeBench v6

Competitive coding, LiveCodeBench v6 • Self-reported

91.9%

Agents' Last Exam

Frontier agentic tasks; Pass@1: 24.3, Score: 51.2 • Self-reported

51.2%

Toolathlon Verified

Real-world tool use, Pass@1 • Self-reported

73.5%

IFBench

Instruction following • Self-reported

81.3%

Humanity's Last Exam

Multidisciplinary reasoning, judged by GPT-4o • Self-reported

35.9%

AndroidWorld

Self-reported by the model provider • Self-reported

84.5%

CharXiv-R

With code interpreter • Self-reported

90.6%

ClawEval-MM

Average score across three trials • Self-reported

60.4%

CoWorkBench

Self-reported by the model provider • Self-reported

73.9%

ERQA

Self-reported by the model provider • Self-reported

72.3%

Humanity's Last Exam

Judged by GPT-4o • Self-reported

35.9%

Job Bench

Self-reported by the model provider • Self-reported

55.7%

LVBench

Self-reported by the model provider • Self-reported

76.6%

MathVision

With code interpreter • Self-reported

95.7%

NL2Repo

Claude Code harness • Self-reported

48.1%

OSWorld 2.0

Binary completion • Self-reported

19.4%

RealWorldQA

Self-reported by the model provider • Self-reported

88.5%

RecreationBench

Self-reported by the model provider • Self-reported

49.9%

Vision2Web

Claude Code harness, judged by gpt-5.4-2026-03-05 • Self-reported

64.0%

License & Metadata

License

qwen

Announcement Date

August 26, 2026

Last Updated

August 27, 2026

Compare Qwen3.8 Flash Next

All comparisons

vs Qwen3.8-27B vs Qwen3.8 Flash vs Qwen3.8 Max vs Kimi K3 vs Gemini 3.6 Flash vs Muse Glimmer-30B vs GLM-5.3 Flash vs GPT-5.6 Sol

Articles about Qwen3.8 Flash Next