Qwen3.5 27B (Reasoning)

AlibabaQwenOpen WeightApache 2.0 · Commercial OK

描述

Qwen3.5-27B is a multimodal dense foundation model with 27 billion parameters. It combines strong reasoning, coding, multilingual, long-context, and visual understanding performance in a production-friendly open-weight package with a native 262K context window.

發布日期

2026-02-24

參數規模

27.0B

上下文長度

262K

支援模態

image, text, video

能力雷達圖

general

coding

reasoning

science估算

agents

multimodal

Science 在缺少專門科學評測時使用推理能力代理估算。

排行榜排名

領域	#排名	分數	來源
智能体与工具	51	57.0	LS
代码能力榜	84	65.0	AA
通用能力榜	43	80.0	AA
多模态榜	60	70.0	LS
推理能力	54	67.0	LS
科学能力	63	68.0	AA

基準測試分數 (LLM Stats)

3d

SUNRGBD

0.35 / 100自報

Hypersim

0.13 / 100自報

Agents

t2-bench

79.0%自報

BFCL-V4

68.5%自報

AndroidWorld_SR

64.2%自報

WideSearch

61.1%自報

BrowseComp

61.0%自報

FullStackBench en

60.1%自報

TIR-Bench

59.8%自報

FullStackBench zh

57.4%自報

OSWorld-Verified

56.2%自報

VITA-Bench

41.9%自報

Terminal-Bench 2.0

41.6%自報

DeepPlanning

22.6%自報

Biology

GPQA

85.5%自報

Chemistry

SuperGPQA

65.6%自報

Code

SWE-Bench Verified

72.4%自報

Communication

Multi-Challenge

60.8%自報

Embodied

EmbSpatialBench

0.84 / 100自報

Finance

MMLU-Pro

86.1%自報

MMLU-ProX

82.2%自報

General

IFEval

95.0%自報

MMLU-Redux

93.2%自報

C-Eval

90.5%自報

MAXIFE

88.0%自報

Global PIQA

87.5%自報

MMMLU

85.9%自報

MMMU

82.3%自報

Include

81.6%自報

MMStar

81.0%自報

LiveCodeBench v6

80.7%自報

IFBench

76.5%自報

MMMU-Pro

75.0%自報

LongBench v2

60.6%自報

NOVA-63

58.1%自報

SimpleVQA

0.56 / 100自報

Grounding

RefCOCO-avg

0.91 / 100自報

ScreenSpot Pro

70.3%自報

RefSpatialBench

0.68 / 100自報

Healthcare

VideoMMMU

82.3%自報

SlakeVQA

80.0%自報

MedXpertQA

62.4%自報

PMC-VQA

62.4%自報

Image To Text

OCRBench

89.4%自報

Language

LingoQA

82.0%自報

WMT24++

77.6%自報

Long Context

MLVU

85.9%自報

LVBench

73.6%自報

AA-LCR

66.1%自報

MMLongBench-Doc

0.60 / 100自報

Math

HMMT 2025

92.0%自報

HMMT25

89.8%自報

MathVista-Mini

87.8%自報

DynaMath

87.7%自報

MathVision

86.0%自報

CodeForces

0.81 / 3000自報

PolyMATH

71.2%自報

Humanity's Last Exam

48.5%自報

Multimodal

VLMsAreBlind

96.9%自報

93.7%自報

AI2D

92.9%自報

MMBench-V1.1

92.6%自報

OmniDocBench 1.5

88.9%自報

VideoMME w sub.

87.0%自報

VideoMME w/o sub.

82.8%自報

CC-OCR

81.0%自報

CharXiv-R

79.5%自報

MVBench

74.6%自報

MMVU

73.3%自報

BabyVision

44.6%自報

ZEROBench-Sub

0.36 / 100自報

Nuscene

15.2%自報

ZEROBench

0.10 / 100自報

Reasoning

CountBench

0.98 / 100自報

Hallusion Bench

70.0%自報

BrowseComp-zh

62.1%自報

ERQA

60.5%自報

Seal-0

47.2%自報

OJBench

40.1%自報

Spatial Reasoning

RealWorldQA

83.7%自報

Vision

ODinW

41.1%自報

AA 評測指數

Intelligence Index

42.1

Coding Index

34.9

Tau2

0.9

Gpqa

0.9

Ifbench

0.8

Lcr

0.7

Scicode

0.4

Terminalbench Hard

0.3

Hle

0.2

LLM Stats 分類評分

Biology

Instruction Following

Structured Output

Text-to-image

Video

Chemistry

Embodied

Finance

General

Grounding

Image To Text

Language

Legal

Math

Physics

Spatial Reasoning

Vision

Economics

Frontend Development

Healthcare

Long Context

Multimodal

Reasoning

Tool Calling

Agents

Code

Communication

Spatial

定價

輸入價格$0.3 / 1M tokens

輸出價格$2.4 / 1M tokens

混合價格(3:1)$0.825 / 1M tokens

速度

Tokens/秒87.6 tokens/s

首Token延遲1.40s

首回答延遲24.23s

可用提供商

(LS 內部計價單位)

提供商	輸入價格	輸出價格
Novita	300K	2.4M

外部連結

LLM Stats