Qwen3.5-4B

Qwen3.5-4B.

Qwen Released February 24th, 2026 Compare with others

Benchmarks with a score

Ranked first, out of five or more models

In the top three

Qwen Feb 2026

Qwen3.5-4B

Open weights

136.5 #211 of 404

Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.

MMMU-Pro 66.3% #58 /90
IFBench 59.2% #56 /80
AA-LCR (Artificial Analysis Long Context Reasoning) 57% #26 /38

Where it ranks

Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.

Knowledge and language

Benchmark	Score	Rank	Source
IFEval	89.8%	#32 of 135	Lab model card
HallusionBench	65%	#12 of 31	Lab model card
AgentIF (2026-10 setting)	62.3%	#12 of 24	Official leaderboard
MMLU-Pro	79.1%	#85 of 167	Lab model card
IFMTBench	66.7%	#5 of 8	Official leaderboard
QIMMA	47.5%	#16 of 25	Official leaderboard
MMLU-Redux	88.8%	#29 of 45	Lab model card
SEA-HELM	63.8	#46 of 71	Official leaderboard
KINA	28.5%	#24 of 37	Official leaderboard
Global PIQA	62.1%	#29 of 44	Official leaderboard
INCLUDE	71%	#21 of 31	Lab model card
AA-LCR (Artificial Analysis Long Context Reasoning)	57%	#26 of 38	Lab model card
LongBench v2	50%	#24 of 35	Lab model card
C-Eval	85.1%	#32 of 46	Lab model card
IFBench	59.2%	#56 of 80	Lab model card
EQ-Bench 3	1,092	#55 of 78	Official leaderboard
BOUQuET	24.7	#13 of 18	Official leaderboard
MMMLU	76.1%	#52 of 71	Lab model card
MMLU-ProX	71.5%	#17 of 23	Lab model card
SuperGPQA	52.9%	#28 of 35	Lab model card
IFStruct	36.3%	#8 of 10	Official leaderboard
MAXIFE	78%	#9 of 11	Lab model card
NOVA-63	54.3%	#9 of 11	Lab model card
SimpleVQA	43.4%	#27 of 30	Lab model card
Wer wird Millionär? LLM Benchmark	77.0	#72 of 77	Official leaderboard
MultiChallenge	25.1%	#40 of 42	Official leaderboard

Math and science

Benchmark	Score	Rank	Source
GPQA Diamond	76.2%	#141 of 302	Lab model card
OTIS Mock AIME 2024-2025	55.8%	#130 of 212	Epoch AI run
PolyMath	51.1%	#15 of 24	Lab model card
HMMT 2025	74%	#42 of 54	Lab model card
HMMT Nov 2025	76.8%	#25 of 28	Lab model card

Coding

Benchmark	Score	Rank	Source
LiveCodeBench v6	55.8%	#77 of 90	Lab model card
OJBench	24.1%	#15 of 16	Lab model card
Aspire	18.6%	#2 of 2	Official leaderboard

Agents and tools

Benchmark	Score	Rank	Source
tau2-bench	79.9%	#38 of 79	Lab model card
ScreenSpot-Pro	60.3%	#24 of 32	Lab model card
AndroidWorld	58.6%	#18 of 23	Lab model card
BFCL V4	50.3%	#26 of 30	Lab model card
LLM Chess	-558	#111 of 123	Official leaderboard
OSWorld-Verified	35.6%	#40 of 41	Lab model card
VITA-Bench	22%	#16 of 16	Lab model card
DeepPlanning	17.6%	#10 of 10	Lab model card

Professional work

Benchmark	Score	Rank	Source
MMEnterprise-Bench	43.8%	#6 of 9	Official leaderboard
MedXpertQA-MM	42.9%	#15 of 21	Lab model card
PMC-VQA	55.5%	#8 of 10	Lab model card
SLAKE	76.1%	#8 of 10	Lab model card

Vision, audio and robotics

Benchmark	Score	Rank	Source
MathVista	85.1%	#17 of 70	Lab model card
MMMU	77.6%	#32 of 103	Lab model card
CountBench	96.3%	#10 of 28	Lab model card
MMStar	78.3%	#13 of 36	Lab model card
AI2D	89.6%	#19 of 52	Lab model card
We-Math	75.4%	#5 of 13	Lab model card
DynaMath	83.3%	#9 of 23	Lab model card
Video-MME	83.5%	#24 of 51	Lab model card
RealWorldQA	79.5%	#18 of 38	Lab model card
MMBench-EN v1.1	89.4%	#11 of 23	Lab model card
ZeroBench	26.3%	#15 of 31	Lab model card
MATH-Vision (MathVision)	74.6%	#27 of 51	Lab model card
VLMs Are Blind	92.6%	#8 of 14	Lab model card
VideoMME v2 (with subtitles)	20.7%	#18 of 31	Official leaderboard
MLVU	82.8%	#13 of 22	Lab model card
RefCOCO / RefCOCO+ / RefCOCOg	88.1%	#18 of 29	Lab model card
ExtractBench	82.4%	#17 of 27	Official leaderboard
GeoGuesser environment (FineEnvs)	0.33	#9 of 14	Official leaderboard
MMMU-Pro	66.3%	#58 of 90	Lab model card
OVO-S-Bench	45.4%	#10 of 15	Official leaderboard
Hypersim	12.5	#6 of 9	Lab model card
LVBench	66.4%	#21 of 31	Lab model card
MMLongBench-Doc	54.2%	#13 of 19	Lab model card
RefSpatial-Bench	54.6%	#11 of 16	Lab model card
ERQA	54%	#25 of 36	Lab model card
OCRBench	85%	#24 of 34	Lab model card
MVBench	71.2%	#17 of 23	Lab model card
CharXiv	70.8%	#56 of 75	Lab model card
EmbSpatialBench	81.3%	#12 of 16	Lab model card
Nuscene	9.90	#6 of 8	Lab model card
Video-MMMU	74.1%	#28 of 36	Lab model card
LingoQA	74.4%	#8 of 10	Lab model card
MMVU	64.9%	#21 of 26	Lab model card
V*	84.3%	#14 of 17	Lab model card
CC-OCR	76.7%	#12 of 14	Lab model card
ParseBench	30.9%	#41 of 44	Official leaderboard
BabyVision	16%	#26 of 26	Lab model card
TIR-Bench	38.9%	#6 of 6	Lab model card

Safety and security

Benchmark	Score		Rank	Source
nocot-bench (NCRI)	44.3		#172 of 185	Official leaderboard

Its lab's numbers put it at 136.5, number 211 of 404

The Self-Reported Index fits one score to the 61 capability results Qwen reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 135.4 to 137.6. How we compute it.

CountBench 96.3
VLMs Are Blind 92.6
IFEval 89.8
AI2D 89.6
MMBench-EN v1.1 89.4
MMLU-Redux 88.8
RefCOCO / RefCOCO+ / RefCOCOg 88.1
C-Eval 85.1
MathVista 85.1
OCRBench 85
V* 84.3
Video-MME 83.5
DynaMath 83.3
MLVU 82.8
EmbSpatialBench 81.3
tau2-bench 79.9
RealWorldQA 79.5
MMLU-Pro 79.1
Global PIQA 78.9
MMStar 78.3
MAXIFE 78
MMMU 77.6
HMMT Nov 2025 76.8
CC-OCR 76.7
GPQA Diamond 76.2
MMMLU 76.1
SLAKE 76.1
We-Math 75.4
MATH-Vision (MathVision) 74.6
LingoQA 74.4
Video-MMMU 74.1
HMMT 2025 74
MMLU-ProX 71.5
MVBench 71.2
INCLUDE 71
CharXiv 70.8
LVBench 66.4
MMMU-Pro 66.3
HallusionBench 65
MMVU 64.9
ScreenSpot-Pro 60.3
IFBench 59.2
AndroidWorld 58.6
AA-LCR (Artificial Analysis Long Context Reasoning) 57
LiveCodeBench v6 55.8
PMC-VQA 55.5
NOVA-63 54.3
MMLongBench-Doc 54.2
SuperGPQA 52.9
PolyMath 51.1
BFCL V4 50.3
LongBench v2 50
MultiChallenge 49
SimpleVQA 43.4
MedXpertQA-MM 42.9
OSWorld-Verified 35.6
ZeroBench 26.3
OJBench 24.1
VITA-Bench 22
DeepPlanning 17.6
BabyVision 16

Its lab's numbers next to 3 leaderboards

Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.

Benchmark	Lab says	Board says	Gap
MultiChallenge	49% MultiChallenge	25.1% Leaderboard, thinking	+24 pts
Global PIQA	78.9% Global PIQA	62.1% Leaderboard, exact_match,strict_match; instruct	+17 pts
MMLU-Pro	79.1% MMLU-Pro	79.1% Leaderboard	Same

Its launch post

The benchmarks the lab chose to headline, with the scores it gave.

Qwen3.5-4B model card (Hugging Face), Qwen3.5 small series Mar 2, 2026

MMLU-Pro 79.1
MMLU-Redux 88.8
C-Eval 85.1
SuperGPQA 52.9
GPQA Diamond 76.2
IFEval 89.8
IFBench 59.2
MultiChallenge 49
AA-LCR 57
LongBench v2 50
HMMT Feb 25 74
HMMT Nov 25 76.8
LiveCodeBench v6 55.8
OJBench 24.1
BFCL-V4 50.3
TAU2-Bench 79.9
VITA-Bench 22
DeepPlanning 17.6
MMMLU 76.1
MMLU-ProX 71.5
NOVA-63 54.3
INCLUDE 71
Global PIQA 78.9
PolyMATH 51.1
WMT24++ 66.6
MAXIFE 78
MMMU 77.6
MMMU-Pro 66.3
MathVision 74.6
Mathvista(mini) 85.1
We-Math 75.4
DynaMath 83.3
Best of ZEROBench / ZEROBench_sub 26.3 26.3
VlmsAreBlind 92.6
BabyVision 16.0/19.1 (with CI / without CI) 16
RealWorldQA 79.5
MMStar 78.3
MMBench EN-DEV-v1.1 89.4
SimpleVQA 43.4
HallusionBench 65
OmniDocBench1.5 86.2
CharXiv(RQ) 70.8
MMLongBench-Doc 54.2
CC-OCR 76.7
AI2D_TEST 89.6
OCRBench 85
ERQA 54
CountBench 96.3
RefCOCO(avg) 88.1
EmbSpatialBench 81.3
RefSpatialBench 54.6
LingoQA 74.4
Hypersim 12.5
Nuscene 9.9
Best of VideoMME(w sub.) / VideoMME(w/o sub.) 83.5 83.5
VideoMMMU 74.1
MLVU 82.8
MVBench 71.2
LVBench 66.4
MMVU 64.9
ScreenSpot Pro 60.3
OSWorld-Verified 35.6
AndroidWorld 58.6
TIR-Bench 38.9/29.9 (with CI / without CI) 38.9
V* 84.3/86.4 (with CI / without CI) 84.3
SLAKE 76.1
PMC-VQA 55.5
MedXpertQA-MM 42.9