跳轉到主要內容

Qwen3.8 27B (xhigh)

AlibabaQwen開源權重Apache 2.0 · 商用許可

描述

Qwen3.8-27B is a dense 27.78-billion-parameter multimodal foundation model for coding, professional work, research, and long-horizon agents. It natively understands text, images, and video, supports configurable thinking effort and preserved reasoning across turns, and uses a 64-layer hybrid architecture with repeating Gated DeltaNet and full-attention blocks. Its native context window is 262,144 tokens and can be extended to approximately 1 million tokens. The open weights are released under Apache 2.0 and support Transformers, vLLM, SGLang, and TokenSpeed.

發布日期
2026-08-14
參數規模
27.8B
上下文長度
1.0M
支援模態
image, text, video

能力雷達圖

34
general
65
coding
91
reasoning
64
science
60
agents
70
multimodal

排行榜排名

領域#排名分數來源
智慧體能力模型榜63
45.0
LS
程式碼能力榜77
86.0
AA
通用能力榜173
56.0
AA
多模態榜6
72.0
LS
科學能力105
71.0
AA

基準測試分數 (LLM Stats)

(LLM Stats (zeroeval))

Agents

AndroidWorld81.9%自報
CoWorkBench70.7%自報
WebArena-Verified64.8%自報
ClawEval-MM56.9%自報
RecreationBench47.1%自報
Agents' Last Exam42.9%自報
Job Bench33.4%自報

Code

QwenSWEBench79.0%自報
Vision2Web62.9%自報
NL2Repo42.3%自報
DeepSWE 1.142.2%自報
SWE-MM38.6%自報

Instruction Following

IFBench79.5%自報

Math

MathVision94.6%自報

Multimodal

OSWorld-Verified84.3%自報

Reasoning

LiveCodeBench v690.3%自報
CharXiv-R90.2%自報
GPQANYU + Cohere + Anthropic (2023)89.2%自報
Terminal-Bench 2.173.0%自報
SWE-Bench ProPrinceton NLP (2024)61.7%自報
SWE-Bench Multimodal38.6%自報
Humanity's Last Exam30.8%自報

Vision

OmniDocBench 1.591.1%自報
RealWorldQA85.9%自報
BabyVision85.6%自報
ERQA65.5%自報

AA 評測指數

(Artificial Analysis)
Gpqa(NYU + Cohere + Anthropic (2023))
90.5
Lcr(Artificial Analysis)
82.0
Terminalbench V2 1
79.8
Coding Index(Artificial Analysis)
68.1
Tau Banking
48.0
Scicode(UIUC + Argonne National Lab (2024))
46.6
Hle(Center for AI Safety + Scale AI (2025))
33.9
Intelligence Index(Artificial Analysis)
33.7

LLM Stats 分類評分

(LLM Stats (zeroeval))
Physics
90
Structured Output
90
Biology
90
Chemistry
90
Instruction Following
80
Spatial Reasoning
80
Multimodal
70
Reasoning
70
General
70
Vision
70
Math
60
Agents
60
Tool Calling
60
Productivity
50
Code
50

定價

輸入價格$0.5 / 1M tokens
輸出價格$3 / 1M tokens
混合價格(3:1)$1.125 / 1M tokens
快取讀取價格$0.085 / 1M tokens

速度

Tokens/秒47.5
首Token延遲1.25s
首回答延遲43.36s

供應商價格排行

供應商價格排行

3 個供應商

最便宜: DeepInfra最貴: Alibaba
供應商輸入輸出
1DeepInfra最便宜
$0
$0
2EmpirioLabs AI
$0.17
$0.5
3Alibaba主要
$0.5
$3

比較該模型在不同 API 供應商之間的定價。

外部連結