Qwen3 VL Thinking (8B) vs Step3 VL 10B
Wins 2 of 7 areas
Agents · Long documents
Wins 4 of 7 areas
Coding · Reasoning · Images and charts · Following instructions
Step3 VL 10B wins more areas, narrowly.Qwen3 VL Thinking (8B) is better at agents.
Scores updated · 34 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 video15Step3 VL 10B5 of 6 tests
- ReasoningHard problems that need careful thinking02Step3 VL 10B2 of 3 tests · 1 tie
- CodingWriting and fixing software02Step3 VL 10B2 of 2 tests
- Following instructionsDoing exactly what it is asked01Step3 VL 10B1 of 1 test
- AgentsCarrying out multi-step tasks on its own10Qwen3 VL Thinking (8B)1 of 1 test
- Long documentsFinding answers in very long texts10Qwen3 VL Thinking (8B)1 of 1 test
- FactsGetting facts right instead of making them up11Even1 each
Agents, 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.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Step3 VL 10B pulls ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+11.1points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+10.3points ahead
- Code for real scientific research problemsSciCode+9.2points ahead
Where Qwen3 VL Thinking (8B) pulls ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+7.1points ahead
- Visual perception tasks people solve at a glanceBLINK+1.9points ahead
Every test, side by side
All 34 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingStep3 VL 10B
- SciCodeStep3 VL 10B by 9.221.931.1+9.2
- Terminal-Bench HardStep3 VL 10B by 1.53.85.3+1.5
ReasoningStep3 VL 10B
- GPQA DiamondStep3 VL 10B by 11.157.969+11.1
- Humanity's Last ExamStep3 VL 10B by 73.810.8+7
- CritPttie0.30tie
FactsEven
- AA-Omniscience · Non-hallucinationStep3 VL 10B by 8.18.616.7+8.1
- AA-Omniscience · AccuracyQwen3 VL Thinking (8B) by 7.120.413.3+7.1
Images and chartsStep3 VL 10B
Long documentsQwen3 VL Thinking (8B)
- AA-LCRQwen3 VL Thinking (8B) by 33.333.30+33.3
Following instructionsStep3 VL 10B
- IFBenchStep3 VL 10B by 10.339.950.2+10.3
Other results18 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.
- HumanEval-VStep3 VL 10B by 39.226.966+39.2
- HMMT 2025Step3 VL 10B by 31.560.692.1+31.5
- LiveCodeBenchStep3 VL 10B by 18.45876.4+18.4
- All-Angles-BenchStep3 VL 10B by 11.345.957.2+11.3
- AA Agentic IndexQwen3 VL Thinking (8B) by 10.215.65.4+10.2
- ReMIStep3 VL 10B by 10.157.267.3+10.1
- MATH-VisionStep3 VL 10B by 8.162.770.8+8.1
- AIME 2025Step3 VL 10B by 7.480.387.7+7.4
- AA-OmniscienceQwen3 VL Thinking (8B) by 6.7-52.3-59+6.7
- τ²-Bench Telecom (AA run)Qwen3 VL Thinking (8B) by 6.422.516.1+6.4
- ScreenSpot-Pro (No tools)Step3 VL 10B by 546.651.5+5
- AI2DStep3 VL 10B by 4.584.989.3+4.5
- Artificial Analysis Coding IndexStep3 VL 10B by 4.19.813.9+4.1
- OSWorld-GStep3 VL 10B by 2.356.759+2.3
- MMBench (EN)Step3 VL 10B by 1.890.592.4+1.8
- PhyXStep3 VL 10B by 1.857.759.5+1.8
- ScreenSpot-V2Qwen3 VL Thinking (8B) by 193.692.6+1
- AA Intelligencetie8.27.7tie
Questions people ask
Which is better, Qwen3 VL Thinking (8B) or Step3 VL 10B?
Step3 VL 10B wins four of the seven areas where both have results: coding, reasoning, images and charts and following instructions. Qwen3 VL Thinking (8B) wins agents and long documents. They are level on facts.
Which is better for coding?
Step3 VL 10B. It wins 2 of the 2 coding tests both models report; Qwen3 VL Thinking (8B) wins none.
How do you compare the two?
We use the 34 benchmark tests both models have published scores on. The verdict counts the 16 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 18 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.