DeepSeek-V3 vs Seed Oss 36B Instruct
Wins 2 of 6 areas
Coding · Facts
Wins 3 of 6 areas
Agents · Reasoning · Long documents
Seed Oss 36B Instruct wins more areas, narrowly.DeepSeek-V3 is better at coding.
Scores updated · 24 tests both models report · How we compare
Where each one wins
Tests won in each of the six 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.
- ReasoningHard problems that need careful thinking02Seed Oss 36B Instruct2 of 3 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own01Seed Oss 36B Instruct1 of 1 test
- Long documentsFinding answers in very long texts01Seed Oss 36B Instruct1 of 1 test
- CodingWriting and fixing software21DeepSeek-V32 of 3 tests
- FactsGetting facts right instead of making them up10DeepSeek-V31 of 2 tests · 1 tie
- Following instructionsDoing exactly what it is asked00Even0 each · 1 tie
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.
Seed Oss 36B Instruct costs 32% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Seed Oss 36B Instruct pulls ahead
- Reasons across sets of long documentsAA-LCR+20.6points ahead
- Recent programming contest problemsLiveCodeBench v6+20.5points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+7.1points ahead
Where DeepSeek-V3 pulls ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+8.4points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+7.3points ahead
- Code for real scientific research problemsSciCode+2.5points ahead
Every test, side by side
All 24 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- LiveCodeBench v6Seed Oss 36B Instruct by 20.546.967.4+20.5
- Terminal-Bench HardDeepSeek-V3 by 8.415.26.8+8.4
- SciCodeDeepSeek-V3 by 2.53936.5+2.5
ReasoningSeed Oss 36B Instruct
- GPQA DiamondSeed Oss 36B Instruct by 7.165.572.6+7.1
- Humanity's Last ExamSeed Oss 36B Instruct by 5.24.79.9+5.2
- CritPttie00tie
FactsDeepSeek-V3
- AA-Omniscience · AccuracyDeepSeek-V3 by 7.325.418.2+7.3
- AA-Omniscience · Non-hallucinationtie14.113.6tie
Long documentsSeed Oss 36B Instruct
- AA-LCRSeed Oss 36B Instruct by 20.640.761.3+20.6
Other results13 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 2024Seed Oss 36B Instruct by 52.539.291.7+52.5
- AIR-Bench 2024Seed Oss 36B Instruct by 34.840.875.6+34.8
- AIME 2025Seed Oss 36B Instruct by 33.451.384.7+33.4
- SimpleQADeepSeek-V3 by 15.224.99.7+15.2
- AA-OmniscienceDeepSeek-V3 by 11.8-40.7-52.5+11.8
- MMLU-ProSeed Oss 36B Instruct by 6.875.982.7+6.8
- Artificial Analysis Coding IndexDeepSeek-V3 by 6.32316.7+6.3
- AA IntelligenceSeed Oss 36B Instruct by 2.49.712.1+2.4
- τ²-Bench Telecom (AA run)Seed Oss 36B Instruct by 2.347.149.4+2.3
- SuperGPQASeed Oss 36B Instruct by 253.755.7+2
- MMLUDeepSeek-V3 by 1.188.587.4+1.1
- MMMLUDeepSeek-V3 by 179.478.4+1
- IFEvaltie86.185.8tie
Questions people ask
Which is better, DeepSeek-V3 or Seed Oss 36B Instruct?
Seed Oss 36B Instruct wins three of the six areas where both have results: agents, reasoning and long documents. DeepSeek-V3 wins coding and facts. They are level on following instructions.
Which is better for coding?
DeepSeek-V3. It wins 2 of the 3 coding tests both models report; Seed Oss 36B Instruct wins 1.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; Seed Oss 36B Instruct costs $0.21 and $0.57. That makes Seed Oss 36B Instruct about 32% cheaper for the same work.
How do you compare the two?
We use the 24 benchmark tests both models have published scores on. The verdict counts the 11 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 13 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.