Enter opens the highlighted result. Esc closes.

Qwen2.5 72B Instruct vs Qwen3 30B A3B

Qwen2.5 72B Instruct leads Qwen3 30B A3B on the Artificial Analysis Intelligence Index, 7.7 to 7.6, and Qwen2.5 72B Instruct costs 1.3x less per output token ($0.40 vs $0.50 per million). Updated October 22, 2026; 43 benchmarks have a result for both.

Replace Qwen2.5 72B Instruct Replace Qwen3 30B A3B
Qwen2.5 72B InstructQwen
Coding
No data
Overall
26.5
Output $/M
$0.40
3
Qwen3 30B A3BQwen
Coding
No data
Overall
38.2
Output $/M
$0.50
40
vs boards won of 43 shared

Qwen2.5 72B Instruct or Qwen3 30B A3B: which to pick

Neither model has a coding score yet, so the two cannot be ranked on coding.

Pick Qwen2.5 72B Instruct for output tokens 1.3x cheaper ($0.40 vs $0.50 per million) and work measured by LiveCodeBench (Artificial Analysis run) or Terminal-Bench Hard (Artificial Analysis run), where it leads.

Model AQwen2.5 72B InstructQwenNot benchmarkedNo coding benchmarkNo cost per solved taskModel BQwen3 30B A3BQwenNot benchmarkedNo coding benchmarkNo cost per solved task
Shared results
43
Qwen2.5 72B Instruct only
10
Qwen3 30B A3B only
9
Like-for-like categories
7 of 9

Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.

Which one for your work

Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.

At a glance

Coding index
Qwen2.5 72B Instruct not benchmarked, Qwen3 30B A3B not benchmarked.
Widest category
Math, 39.1 points apart on average over 5 shared benchmarks.
Widest single gap
AIME 2025 (Artificial Analysis run): Qwen2.5 72B Instruct 14.0% against Qwen3 30B A3B 72.3%, 58.3 points apart on a 0 to 100 scale.
Output price
Qwen2.5 72B Instruct $0.40 against Qwen3 30B A3B $0.50 per million output tokens, 1.3x apart.
Context window
Qwen2.5 72B Instruct takes 33K tokens, Qwen3 30B A3B 41K.

Benchmark by benchmark

First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.

Boards won, where both have a result
CategoryQwen2.5 72B InstructBarsQwen3 30B A3BTies
Coding05
Intelligence11
Agentic work10
Reasoning06
Math05
Knowledge05
General118
All shared boards340
CodingLike-for-likeSix of these boards feed the Vibe Coding Index05
Coding benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
LiveCodeBench (Artificial Analysis run)27.6%50.6%Qwen3 30B A3B
Terminal-Bench Hard (Artificial Analysis run)4.5%6.8%Qwen3 30B A3B
WeirdML16.0%29.8%Qwen3 30B A3B
LMArena Text: coding12921337Qwen3 30B A3B
LMArena Text: software and IT12921340Qwen3 30B A3B
Neither model has a result on the other 34 boards here
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index11
Intelligence benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
AA Intelligence Index7.77.6Qwen2.5 72B Instruct
Epoch Capabilities Index129.0137.3Qwen3 30B A3B
Agentic workLike-for-likePublic boards, shown as published, kept out of the index10
Agentic work benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
τ²-Bench Telecom34.5%26.0%Qwen2.5 72B Instruct
METR time horizons35.8%Not measuredOnly Qwen2.5 72B Instruct measured
BALROG16.2%Not measuredOnly Qwen2.5 72B Instruct measured
TheAgentCompany5.7%Not measuredOnly Qwen2.5 72B Instruct measured
Neither model has a result on the other 37 boards here
ReasoningLike-for-likePublic boards, shown as published, kept out of the index06
Reasoning benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
GPQA Diamond (Artificial Analysis run)49.1%61.6%Qwen3 30B A3B
Humanity's Last Exam (Artificial Analysis run)3.6%6.2%Qwen3 30B A3B
GPQA Diamond (Epoch AI run)49.1%61.7%Qwen3 30B A3B
BIG-Bench Hard79.8%Not measuredOnly Qwen2.5 72B Instruct measured
LMArena Text: hard prompts12711314Qwen3 30B A3B
LMArena Text: expert12451314Qwen3 30B A3B
LMArena Text: hard prompts in English12791329Qwen3 30B A3B
Neither model has a result on the other 15 boards here
MathLike-for-likePublic boards, shown as published, kept out of the index05
Math benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
AA Math Index14.072.3Qwen3 30B A3B
AIME 2025 (Artificial Analysis run)14.0%72.3%Qwen3 30B A3B
OTIS Mock AIME8.1%62.8%Qwen3 30B A3B
MATH level 563.2%Not measuredOnly Qwen2.5 72B Instruct measured
LMArena Text: math12831355Qwen3 30B A3B
LMArena Text: mathematical work12901363Qwen3 30B A3B
Neither model has a result on the other 7 boards here
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index05
Knowledge benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
MMLU Pro (Artificial Analysis run)72.0%77.7%Qwen3 30B A3B
MMLU85.3%Not measuredOnly Qwen2.5 72B Instruct measured
AI2 Reasoning Challenge94.5%Not measuredOnly Qwen2.5 72B Instruct measured
LMArena Text: science12751321Qwen3 30B A3B
LMArena Text: medicine and healthcare12581318Qwen3 30B A3B
LMArena Text: legal and government12881314Qwen3 30B A3B
LMArena Text: business and finance12531309Qwen3 30B A3B
Neither model has a result on the other 11 boards here
Web designDirectional onlyPublic boards, shown as published, kept out of the index5 boards, none shared
Web design benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
Design Arena: websitesNot measured954Only Qwen3 30B A3B measured
Design Arena: UI componentsNot measured943Only Qwen3 30B A3B measured
Design Arena: game devNot measured898Only Qwen3 30B A3B measured
Design Arena: data vizNot measured967Only Qwen3 30B A3B measured
Design Arena: all code categoriesNot measured941Only Qwen3 30B A3B measured
Neither model has a result on the other 4 boards here
GeneralLike-for-likePublic boards, shown as published, kept out of the index118
General benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
IFBench36.9%41.5%Qwen3 30B A3B
Long context reasoning (AA-LCR)Not measured0.0%Only Qwen3 30B A3B measured
DTBench62.9%60.3%Qwen2.5 72B Instruct
LMCA13.4%15.8%Qwen3 30B A3B
Fiction.LiveBenchNot measured40.6%Only Qwen3 30B A3B measured
Creative writing (Lech Mazur)Not measured75.3%Only Qwen3 30B A3B measured
HellaSwag84.8%Not measuredOnly Qwen2.5 72B Instruct measured
WinoGrande82.3%Not measuredOnly Qwen2.5 72B Instruct measured
PIQA82.6%Not measuredOnly Qwen2.5 72B Instruct measured
LMArena Text12691317Qwen3 30B A3B
LMArena Text: instruction following12541283Qwen3 30B A3B
LMArena Text: creative writing12211271Qwen3 30B A3B
LMArena Text: English12831335Qwen3 30B A3B
LMArena Text: non-English12521295Qwen3 30B A3B
LMArena Text: Chinese12731347Qwen3 30B A3B
LMArena Text: German12341307Qwen3 30B A3B
LMArena Text: Japanese11801254Qwen3 30B A3B
LMArena Text: Korean11881261Qwen3 30B A3B
LMArena Text: French12801352Qwen3 30B A3B
LMArena Text: Spanish12551317Qwen3 30B A3B
LMArena Text: Russian12641291Qwen3 30B A3B
LMArena Text: PolishNot measured1299Only Qwen3 30B A3B measured
LMArena Text: multi-turn12721307Qwen3 30B A3B
LMArena Text: longer queries12811312Qwen3 30B A3B
LMArena Text: writing and language12471290Qwen3 30B A3B
LMArena Text: entertainment and media12111258Qwen3 30B A3B
Neither model has a result on the other 20 boards here
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
Image generation benchmarks for Qwen2.5 72B Instruct and Qwen3 30B A3B
BenchmarkQwen2.5 72B InstructQwen3 30B A3BLeader
Neither model has a result on 5 boards here

What each workload costs

Three fixed token mixes at each model's listed price, so per-token rates become a bill.

Chat turn

1K input + 500 output tokens

Qwen2.5 72B Instruct$0.0006Fits in one request
Qwen3 30B A3B$0.0004Fits in one request

Qwen3 30B A3B costs less

Repository review

50K input + 3K output tokens

Qwen2.5 72B Instruct$0.019Exceeds the context window
Qwen3 30B A3B$0.0075Exceeds the context window

Qwen3 30B A3B costs less

Agent loop

200K input + 10K output tokens

Qwen2.5 72B Instruct$0.076Exceeds the context window
Qwen3 30B A3B$0.029Exceeds the context window

Qwen3 30B A3B costs less

Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.

Specification differences

The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.

Specifications of Qwen2.5 72B Instruct and Qwen3 30B A3B
SpecificationQwen2.5 72B InstructQwen3 30B A3B
MakerQwenQwen
WeightsOpen weightsOpen weights
ReleasedSeptember 19, 2024April 28, 2025
Context window33K41K
Input $/Mtok$0.36$0.12
Output $/Mtok$0.40$0.50
Coding indexNot benchmarkedNot benchmarked
Independent measurements behind it00
Cost per solved taskNot computableNot computable
SourceOpenRouter listingOpenRouter listing

Questions

Is Qwen2.5 72B Instruct better than Qwen3 30B A3B for coding?

Neither model has a coding score yet, so the two cannot be ranked on coding. Pick Qwen2.5 72B Instruct for output tokens 1.3x cheaper ($0.40 vs $0.50 per million) and work measured by LiveCodeBench (Artificial Analysis run) or Terminal-Bench Hard (Artificial Analysis run), where it leads.

Which is cheaper, Qwen2.5 72B Instruct or Qwen3 30B A3B?

Qwen2.5 72B Instruct costs $0.36 per million input tokens and $0.40 per million output tokens; Qwen3 30B A3B costs $0.12 and $0.50. Cost per solved task is known for neither of them.

Which has the bigger context window, Qwen2.5 72B Instruct or Qwen3 30B A3B?

Qwen3 30B A3B, with 41K tokens against 33K.

How many benchmarks is the Qwen2.5 72B Instruct vs Qwen3 30B A3B comparison based on?

43 benchmarks have a result for both models as of October 22, 2026. Qwen2.5 72B Instruct's coding index rests on 0 independent benchmarks and Qwen3 30B A3B's on 0. Where only one model has a result, the page says so instead of naming a winner.

Compare next