Qwen3.5 122B A10B vs Qwen3.5 27B
Wins 3 of 7 areas
Agents · Reasoning · Images and charts
Wins 2 of 7 areas
Coding · Long documents
Qwen3.5 122B A10B wins more areas, narrowly.Qwen3.5 27B is cheaper and better at coding.
Both rank among the ten best models we track in images and charts.
Scores updated · 74 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 video31Qwen3.5 122B A10B3 of 7 tests · 3 ties
- AgentsCarrying out multi-step tasks on its own21Qwen3.5 122B A10B2 of 3 tests
- ReasoningHard problems that need careful thinking10Qwen3.5 122B A10B1 of 3 tests · 2 ties
- CodingWriting and fixing software04Qwen3.5 27B4 of 7 tests · 3 ties
- Long documentsFinding answers in very long texts01Qwen3.5 27B1 of 2 tests · 1 tie
- FactsGetting facts right instead of making them up11Even1 each · 1 tie
- Following instructionsDoing exactly what it is asked00Even0 each · 2 ties
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.
Qwen3.5 27B costs 25% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Qwen3.5 122B A10B pulls ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+3.7points ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+2.8points ahead
- Reads text in imagesOCRBench+2.7points ahead
Where Qwen3.5 27B pulls ahead
- Real work tasks from 44 professionsGDPVal+17.3points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+7.3points ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+6.3points ahead
Every test, side by side
All 74 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingQwen3.5 27B
- SWE-bench MultilingualQwen3.5 27B by 7.36269.3+7.3
- SWE-bench ProQwen3.5 27B by 6.344.951.2+6.3
- LiveCodeBench v6Qwen3.5 27B by 1.878.980.7+1.8
- Terminal-Bench HardQwen3.5 27B by 1.531.132.6+1.5
- SWE-bench Verifiedtie7272.4tie
- LMArena · WebDevtie13601357tie
- SciCodetie39.739.5tie
AgentsQwen3.5 122B A10B
- GDPValQwen3.5 27B by 17.315.733+17.3
- BrowseCompQwen3.5 122B A10B by 2.863.861+2.8
- OSWorld-VerifiedQwen3.5 122B A10B by 1.85856.2+1.8
ReasoningQwen3.5 122B A10B
- Humanity's Last ExamQwen3.5 122B A10B by 1.325.223.9+1.3
- CritPttie0.60.9tie
- GPQA Diamondtie85.785.8tie
FactsEven
- AA-Omniscience · Non-hallucinationQwen3.5 27B by 5.612.918.5+5.6
- AA-Omniscience · AccuracyQwen3.5 122B A10B by 3.724.420.7+3.7
- Vectara HHEM hallucination ratelower is bettertie11.212.1tie
Images and chartsQwen3.5 122B A10B
- OCRBenchQwen3.5 122B A10B by 2.792.189.4+2.7
- CharXiv (RQ)Qwen3.5 27B by 2.377.279.5+2.3
- MMStarQwen3.5 122B A10B by 1.982.981+1.9
- MMMUQwen3.5 122B A10B by 1.683.982.3+1.6
- LMArena · Visiontie12451241tie
- MathVistatie87.487.8tie
- MMMU-Protie7575tie
Long documentsQwen3.5 27B
- AA-LCRQwen3.5 27B by 1.476.377.7+1.4
- LongBench v2tie60.260.6tie
Following instructionsEven
- Multi-Challengetie61.560.8tie
- IFBenchtie75.775.6tie
Other results47 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.
- AA Agentic IndexQwen3.5 27B by 459.654.6+45
- Artificial Analysis Coding IndexQwen3.5 122B A10B by 10.845.734.9+10.8
- VITA-BenchQwen3.5 27B by 8.333.641.9+8.3
- vectara_avg_summary_lengthQwen3.5 27B by 886.494.4+8
- BrowseComp-ZHQwen3.5 122B A10B by 7.869.962.1+7.8
- Terminal-Bench 2.0Qwen3.5 122B A10B by 7.849.441.6+7.8
- AA IntelligenceQwen3.5 27B by 7.315.622.9+7.3
- SimpleVQAQwen3.5 122B A10B by 5.761.756+5.7
- BabyVisionQwen3.5 27B by 4.440.244.6+4.4
- BFCLv4Qwen3.5 122B A10B by 3.772.268.5+3.7
- Seal-0Qwen3.5 27B by 3.144.147.2+3.1
- AA-OmniscienceQwen3.5 122B A10B by 2.5-41.5-44+2.5
- HallusionBenchQwen3.5 27B by 2.467.670+2.4
- MV-BenchQwen3.5 122B A10B by 276.674.6+2
- DynaMathQwen3.5 27B by 1.885.987.7+1.8
- IFEvalQwen3.5 27B by 1.693.495+1.6
- RefSpatial-BenchQwen3.5 122B A10B by 1.669.367.7+1.6
- DeepPlanningQwen3.5 122B A10B by 1.524.122.6+1.5
- ERQAQwen3.5 122B A10B by 1.56260.5+1.5
- SuperGPQAQwen3.5 122B A10B by 1.567.165.6+1.5
- C-EvalQwen3.5 122B A10B by 1.491.990.5+1.4
- MLVUQwen3.5 122B A10B by 1.487.385.9+1.4
- MMVUQwen3.5 122B A10B by 1.474.773.3+1.4
- RealWorldQAQwen3.5 122B A10B by 1.485.183.7+1.4
- IncludeQwen3.5 122B A10B by 1.282.881.6+1.2
- MMlongBench DocQwen3.5 27B by 1.25960.2+1.2
- ZeroBenchQwen3.5 27B by 1910+1
- OmniDocBench 1.5tie89.888.9tie
- vectara_factual_consistencytie88.887.9tie
- CC-OCRtie81.881tie
- LVBenchtie74.473.6tie
- MMLU-Reduxtie9493.2tie
- MMMLUtie86.785.9tie
- WMT24++tie78.377.6tie
- EmbSpatial-Benchtie83.984.5tie
- MMLU-Protie86.786.1tie
- OJBenchtie39.540.1tie
- WideSearchtie60.561.1tie
- HMMT 2025tie90.389.8tie
- t2-benchtie79.579tie
- AI2Dtie93.392.9tie
- VideoMMMUtie8282.3tie
- τ²-Bench Telecom (AA run)tie93.693.9tie
- MATH-Visiontie86.286tie
- ScreenSpot-Pro (No tools)tie70.470.3tie
- MMLU-ProXtie82.282.2tie
- vectara_answer_ratetie99.899.8tie
Questions people ask
Which is better, Qwen3.5 122B A10B or Qwen3.5 27B?
Qwen3.5 122B A10B wins three of the seven areas where both have results: agents, reasoning and images and charts. Qwen3.5 27B wins coding and long documents, and costs 25% less. They are level on facts and following instructions.
Which is better for coding?
Qwen3.5 27B. It wins 4 of the 7 coding tests both models report; Qwen3.5 122B A10B wins none, and 3 are ties.
Which is cheaper?
Qwen3.5 122B A10B costs $0.40 per million input tokens and $3.20 per million output tokens; Qwen3.5 27B costs $0.30 and $2.40. That makes Qwen3.5 27B about 25% cheaper for the same work.
How do you compare the two?
We use the 74 benchmark tests both models have published scores on. The verdict counts the 27 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 47 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.