DeepSeek-V3.2 vs Gemma 4 31B
Wins 6 of 7 areas
Coding · Agents · Reasoning · Facts · Math · Long documents
Wins 1 of 7 areas
Following instructions
DeepSeek-V3.2 is the stronger all-rounder.Gemma 4 31B is cheaper and better at following instructions.
Scores updated · 43 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.
- FactsGetting facts right instead of making them up40DeepSeek-V3.24 of 4 tests
- AgentsCarrying out multi-step tasks on its own30DeepSeek-V3.23 of 3 tests
- MathCompetition and research-level math30DeepSeek-V3.23 of 3 tests
- CodingWriting and fixing software42DeepSeek-V3.24 of 8 tests · 2 ties
- ReasoningHard problems that need careful thinking21DeepSeek-V3.22 of 3 tests
- Long documentsFinding answers in very long texts10DeepSeek-V3.21 of 1 test
- Following instructionsDoing exactly what it is asked01Gemma 4 31B1 of 1 test
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.
Gemma 4 31B costs 19% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V3.2 pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+21.1points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+18.5points ahead
- Short factual questions, answered correctlySimpleQA Verified+17.1points ahead
Where Gemma 4 31B pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+20.1points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+14.9points ahead
- Code for real scientific research problemsSciCode+6.6points ahead
Every test, side by side
All 43 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3.2
- SWE-bench VerifiedDeepSeek-V3.2 by 21.173.152+21.1
- SWE-bench ProGemma 4 31B by 20.115.635.7+20.1
- SWE-bench MultilingualDeepSeek-V3.2 by 18.570.251.7+18.5
- SciCodeGemma 4 31B by 6.638.945.5+6.6
- Terminal-Bench 2.1DeepSeek-V3.2 by 3.446.843.4+3.4
- LiveCodeBench v6DeepSeek-V3.2 by 3.383.380+3.3
- Terminal-Bench Hardtie35.636.4tie
- LMArena · WebDevtie13621365tie
AgentsDeepSeek-V3.2
- MCP AtlasDeepSeek-V3.2 by 562.257.2+5
- τ-Bench V3 · BankingDeepSeek-V3.2 by 418.814.8+4
- GDPValDeepSeek-V3.2 by 3.79.86.1+3.7
ReasoningDeepSeek-V3.2
- GPQA DiamondGemma 4 31B by 1.78485.7+1.7
- CritPtDeepSeek-V3.2 by 1.52.91.4+1.5
- Humanity's Last ExamDeepSeek-V3.2 by 124.623.6+1
FactsDeepSeek-V3.2
- SimpleQA VerifiedDeepSeek-V3.2 by 17.127.510.4+17.1
- AA-Omniscience · AccuracyDeepSeek-V3.2 by 133320+13
- AA-Omniscience · Non-hallucinationDeepSeek-V3.2 by 2.317.315+2.3
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3.2 by 1.16.37.4+1.1
MathDeepSeek-V3.2
- HMMT Feb. 2026DeepSeek-V3.2 by 6.984.177.2+6.9
- AIME 2026DeepSeek-V3.2 by 594.289.2+5
- IMOAnswerBenchDeepSeek-V3.2 by 3.878.374.5+3.8
Following instructionsGemma 4 31B
- IFBenchGemma 4 31B by 14.960.775.6+14.9
Other results20 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.
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 30.790.659.9+30.7
- AA-OmniscienceDeepSeek-V3.2 by 25.4-22.5-47.9+25.4
- MCPMarkDeepSeek-V3.2 by 19.93818.1+19.9
- HLE (with tools)DeepSeek-V3.2 by 14.340.826.5+14.3
- Tool-DecathlonDeepSeek-V3.2 by 1435.221.2+14
- vectara_avg_summary_lengthGemma 4 31B by 13.86275.8+13.8
- AA Agentic IndexDeepSeek-V3.2 by 11.618.36.7+11.6
- vectara_answer_rateGemma 4 31B by 7.492.6100+7.4
- AA IntelligenceDeepSeek-V3.2 by 6.821.514.7+6.8
- t2-benchGemma 4 31B by 6.280.286.4+6.2
- GPQA (unspecified)Gemma 4 31B by 4.479.984.3+4.4
- HMMT Feb. 2025DeepSeek-V3.2 by 3.892.588.7+3.8
- Terminal-Bench 2.0DeepSeek-V3.2 by 3.546.442.9+3.5
- τ²-BenchDeepSeek-V3.2 by 3.480.376.9+3.4
- HMMT Nov. 2025DeepSeek-V3.2 by 2.59087.5+2.5
- τ³-BenchDeepSeek-V3.2 by 1.769.267.5+1.7
- vectara_factual_consistencyDeepSeek-V3.2 by 1.193.792.6+1.1
- Artificial Analysis Coding Indextie44.243.4tie
- MMLU-Protie8585.2tie
- MMLU-Reduxtie93.793.7tie
Questions people ask
Which is better, DeepSeek-V3.2 or Gemma 4 31B?
DeepSeek-V3.2 wins six of the seven areas where both have results: coding, agents, reasoning, facts, math and long documents. Gemma 4 31B wins following instructions, and costs 19% less.
Which is better for coding?
DeepSeek-V3.2. It wins 4 of the 8 coding tests both models report; Gemma 4 31B wins 2, and 2 are ties.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; Gemma 4 31B costs $0.17 and $0.40. That makes Gemma 4 31B about 19% cheaper for the same work.
How do you compare the two?
We use the 43 benchmark tests both models have published scores on. The verdict counts the 23 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 20 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.