跳轉到主要內容

Qwen3.5 4B (Non-reasoning)

AlibabaQwenOpen WeightApache 2.0 · Commercial OK

描述

Qwen3.5-4B is a 4 billion parameter vision-language model using Gated DeltaNet hybrid architecture with a 3:1 ratio of linear attention to full softmax attention. It supports 262K native context length and delivers strong performance for its size across knowledge, reasoning, coding, and multilingual tasks.

發布日期
2026-03-02
參數規模
4.0B
上下文長度
支援模態

能力雷達圖

19
general
14
coding
71
reasoning
40
science估算
70
agents
50
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78
47.0
LS
代码能力榜292
26.0
AA
通用能力榜222
46.0
AA
科学能力281
39.0
AA

基準測試分數 (LLM Stats)

Agents

t2-bench79.9%自報
BFCL-V450.3%自報
VITA-Bench22.0%自報
DeepPlanning17.6%自報

Biology

GPQA76.2%自報

Chemistry

SuperGPQA52.9%自報

Communication

Multi-Challenge49.0%自報

Finance

MMLU-Pro79.1%自報
MMLU-ProX71.5%自報

General

IFEval89.8%自報
MMLU-Redux88.8%自報
C-Eval85.1%自報
Global PIQA78.9%自報
MAXIFE78.0%自報
MMMLU76.1%自報
Include71.0%自報
IFBench59.2%自報
LiveCodeBench v655.8%自報
NOVA-6354.3%自報
LongBench v250.0%自報

Language

WMT24++66.6%自報

Long Context

AA-LCR57.0%自報

Math

HMMT2576.8%自報
HMMT 202574.0%自報
PolyMATH51.1%自報

AA 評測指數

Intelligence Index
22.6
Coding Index
13.7
Tau2
0.9
Gpqa
0.7
Ifbench
0.3
Lcr
0.3
Scicode
0.2
Terminalbench Hard
0.1
Hle
0.1

LLM Stats 分類評分

Biology
80
Language
80
Structured Output
70
Tool Calling
70
Finance
70
General
70
Healthcare
70
Instruction Following
70
Legal
70
Math
70
Physics
70
Chemistry
60
Reasoning
60
Spatial Reasoning
50
Vision
50
Communication
50
Economics
50
Long Context
50
Multimodal
50
Agents
40

定價

輸入價格$0.03 / 1M tokens
輸出價格$0.15 / 1M tokens
混合價格(3:1)$0.06 / 1M tokens

速度

Tokens/秒216.4 tokens/s
首Token延遲0.25s
首回答延遲0.25s

可用提供商

(LS 內部計價單位)

暫無提供商資料

外部連結