Qwen3.5 4B (Non-reasoning)
AlibabaQwen開源權重Apache 2.0 · 商用許可
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
10
general
20
coding
71
reasoning
52
science
70
agents
50
multimodal
排行榜排名
基準測試分數 (LLM Stats)
(LLM Stats (zeroeval))Agents
t2-bench
79.9%自報
VITA-Bench
22.0%自報
DeepPlanning
17.6%自報
Chat
IFEvalGoogle Research (2023)
89.8%自報
Multi-Challenge
49.0%自報
General
C-Eval
85.1%自報
MAXIFE
78.0%自報
Include
71.0%自報
NOVA-63
54.3%自報
Instruction Following
IFBench
59.2%自報
Language
MMLU-Redux
88.8%自報
MMLU-Pro
79.1%自報
MMMLU
76.1%自報
MMLU-ProX
71.5%自報
WMT24++
66.6%自報
Long Context
LongBench v2
50.0%自報
Math
HMMT25
76.8%自報
HMMT 2025
74.0%自報
PolyMATH
51.1%自報
Reasoning
Global PIQA
78.9%自報
GPQANYU + Cohere + Anthropic (2023)
76.2%自報
AA-LCR
57.0%自報
LiveCodeBench v6
55.8%自報
SuperGPQA
52.9%自報
Tool Calling
BFCL-V4
50.3%自報
AA 評測指數
(Artificial Analysis)Tau2(Sierra + U Toronto + Vector Institute (2025))87.7
Gpqa(NYU + Cohere + Anthropic (2023))71.2
Lcr(Artificial Analysis)34.7
Ifbench(Google Research (2023))33.3
Terminalbench V2 121.3
Coding Index(Artificial Analysis)20.3
Terminalbench Hard(Stanford × Laude Institute (2026))11.4
Intelligence Index(Artificial Analysis)10.8
Hle(Center for AI Safety + Scale AI (2025))8.0
Tau Banking4.3
LLM Stats 分類評分
(LLM Stats (zeroeval))Language80
Biology80
Chat70
Instruction Following70
Legal70
Math70
Physics70
Structured Output70
Finance70
Healthcare70
Tool Calling70
Reasoning60
General60
Chemistry60
Long Context50
Multimodal50
Spatial Reasoning50
Communication50
Economics50
Vision50
Agents40
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25.9
首Token延遲0.53s
首回答延遲0.53s
供應商價格排行
供應商價格排行
1 個供應商
供應商輸入輸出
1Alibaba主要
$0.03
$0.15
比較該模型在不同 API 供應商之間的定價。