Qwen3 VL 30B A3B Reasoning vs Qwen3 VL 8B Instruct
Wins 5 of 7 areas
Coding · Reasoning · Images and charts · Long documents · Following instructions
Wins 1 of 7 areas
Agents
Qwen3 VL 30B A3B Reasoning is the stronger all-rounder.Qwen3 VL 8B Instruct is cheaper and better at agents.
Scores updated · 51 tests both models report · How we compare
Where each one wins
Tests won in each of the seven areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- Images and chartsUnderstanding pictures, charts and video52Qwen3 VL 30B A3B Reasoning5 of 7 tests
- CodingWriting and fixing software30Qwen3 VL 30B A3B Reasoning3 of 3 tests
- ReasoningHard problems that need careful thinking20Qwen3 VL 30B A3B Reasoning2 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts10Qwen3 VL 30B A3B Reasoning1 of 1 test
- Following instructionsDoing exactly what it is asked10Qwen3 VL 30B A3B Reasoning1 of 1 test
- AgentsCarrying out multi-step tasks on its own01Qwen3 VL 8B Instruct1 of 2 tests · 1 tie
- FactsGetting facts right instead of making them up11Even1 each
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Qwen3 VL 8B Instruct costs 66% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Qwen3 VL 30B A3B Reasoning pulls ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+29.3points ahead
- Reasons across sets of long documentsAA-LCR+26points ahead
- Recent programming contest problemsLiveCodeBench v6+24.9points ahead
Where Qwen3 VL 8B Instruct pulls ahead
- Reads text in imagesOCRBench+5.7points ahead
- Visual perception tasks people solve at a glanceBLINK+3.7points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+3.6points ahead
Every test, side by side
All 51 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingQwen3 VL 30B A3B Reasoning
- LiveCodeBench v6Qwen3 VL 30B A3B Reasoning by 24.964.239.3+24.9
- SciCodeQwen3 VL 30B A3B Reasoning by 11.428.817.4+11.4
- Terminal-Bench HardQwen3 VL 30B A3B Reasoning by 35.32.3+3
AgentsQwen3 VL 8B Instruct
- OSWorld-VerifiedQwen3 VL 8B Instruct by 3.330.633.9+3.3
- GDPValtie8.39.1tie
ReasoningQwen3 VL 30B A3B Reasoning
- GPQA DiamondQwen3 VL 30B A3B Reasoning by 29.37242.7+29.3
- Humanity's Last ExamQwen3 VL 30B A3B Reasoning by 6.28.92.7+6.2
- CritPttie00tie
FactsEven
- AA-Omniscience · AccuracyQwen3 VL 8B Instruct by 3.616.920.5+3.6
- AA-Omniscience · Non-hallucinationQwen3 VL 30B A3B Reasoning by 1.210.29+1.2
Images and chartsQwen3 VL 30B A3B Reasoning
- MMMU-ProQwen3 VL 30B A3B Reasoning by 14.561.847.3+14.5
- CharXiv (RQ)Qwen3 VL 30B A3B Reasoning by 10.256.646.4+10.2
- OCRBenchQwen3 VL 8B Instruct by 5.783.989.6+5.7
- MathVistaQwen3 VL 30B A3B Reasoning by 4.781.977.2+4.7
- MMStarQwen3 VL 30B A3B Reasoning by 4.675.570.9+4.6
- BLINKQwen3 VL 8B Instruct by 3.765.469.1+3.7
- Video-MMEQwen3 VL 30B A3B Reasoning by 1.973.371.4+1.9
Long documentsQwen3 VL 30B A3B Reasoning
- AA-LCRQwen3 VL 30B A3B Reasoning by 2642.716.7+26
Following instructionsQwen3 VL 30B A3B Reasoning
- IFBenchQwen3 VL 30B A3B Reasoning by 12.845.132.3+12.8
Other results32 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- AIME 2025Qwen3 VL 30B A3B Reasoning by 37.283.145.9+37.2
- HMMT 2025Qwen3 VL 30B A3B Reasoning by 35.167.632.5+35.1
- SuperGPQAQwen3 VL 30B A3B Reasoning by 11.956.444.5+11.9
- MATH-VisionQwen3 VL 30B A3B Reasoning by 11.865.753.9+11.8
- MMLU-ProXQwen3 VL 30B A3B Reasoning by 10.776.165.4+10.7
- VideoMMMUQwen3 VL 30B A3B Reasoning by 9.77565.3+9.7
- τ²-Bench Telecom (AA run)Qwen3 VL 8B Instruct by 9.319.929.2+9.3
- MMLU-ProQwen3 VL 30B A3B Reasoning by 8.980.571.6+8.9
- IncludeQwen3 VL 30B A3B Reasoning by 7.574.567+7.5
- CharadesSTAQwen3 VL 30B A3B Reasoning by 6.762.756+6.7
- MMMU (val) (Pass@1)Qwen3 VL 30B A3B Reasoning by 6.47669.6+6.4
- MMLU-ReduxQwen3 VL 30B A3B Reasoning by 690.984.9+6
- RealWorldQAQwen3 VL 30B A3B Reasoning by 5.977.471.5+5.9
- Artificial Analysis Coding IndexQwen3 VL 30B A3B Reasoning by 5.813.17.3+5.8
- AA-OmniscienceQwen3 VL 8B Instruct by 5.7-57.6-51.9+5.7
- HallusionBenchQwen3 VL 30B A3B Reasoning by 4.96661.1+4.9
- AA Agentic IndexQwen3 VL 8B Instruct by 3.514.918.4+3.5
- MV-BenchQwen3 VL 30B A3B Reasoning by 3.37268.7+3.3
- InfoVQA_testQwen3 VL 30B A3B Reasoning by 2.98683.1+2.9
- OCRBench v2_enQwen3 VL 8B Instruct by 2.862.665.4+2.8
- ScreenSpot-Pro (No tools)Qwen3 VL 30B A3B Reasoning by 2.757.354.6+2.7
- BFCL v3Qwen3 VL 30B A3B Reasoning by 2.368.666.3+2.3
- AA IntelligenceQwen3 VL 30B A3B Reasoning by 2.29.57.3+2.2
- CC-OCRQwen3 VL 8B Instruct by 2.177.879.9+2.1
- IFEvalQwen3 VL 8B Instruct by 281.783.7+2
- AI2DQwen3 VL 30B A3B Reasoning by 1.286.985.7+1.2
- LVBenchQwen3 VL 30B A3B Reasoning by 1.259.258+1.2
- DocVQA_testQwen3 VL 8B Instruct by 1.19596.1+1.1
- OCRBench v2 (Chinese)tie60.461.2tie
- ERQAtie45.345.8tie
- ScreenSpottie94.794.4tie
- MM MTBenchtie7.97.7tie
Questions people ask
Which is better, Qwen3 VL 30B A3B Reasoning or Qwen3 VL 8B Instruct?
Qwen3 VL 30B A3B Reasoning wins five of the seven areas where both have results: coding, reasoning, images and charts, long documents and following instructions. Qwen3 VL 8B Instruct wins agents, and costs 66% less. They are level on facts.
Which is better for coding?
Qwen3 VL 30B A3B Reasoning. It wins 3 of the 3 coding tests both models report; Qwen3 VL 8B Instruct wins none.
Which is cheaper?
Qwen3 VL 30B A3B Reasoning costs $0.20 per million input tokens and $2.40 per million output tokens; Qwen3 VL 8B Instruct costs $0.18 and $0.70. That makes Qwen3 VL 8B Instruct about 66% cheaper for the same work.
How do you compare the two?
We use the 51 benchmark tests both models have published scores on. The verdict counts the 19 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 32 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.