DeepSeek-R1 vs Granite 3.3 8B Instruct
Wins 5 of 6 areas
Coding · Reasoning · Facts · Long documents · Following instructions
Wins 0 of 6 areas
—
DeepSeek-R1 is the stronger all-rounder.Granite 3.3 8B Instruct is cheaper.
Scores updated · 28 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 thinking20DeepSeek-R12 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up20DeepSeek-R12 of 3 tests · 1 tie
- CodingWriting and fixing software20DeepSeek-R12 of 2 tests
- Long documentsFinding answers in very long texts10DeepSeek-R11 of 1 test
- Following instructionsDoing exactly what it is asked10DeepSeek-R11 of 1 test
- AgentsCarrying out multi-step tasks on its own00Even0 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.
Granite 3.3 8B Instruct costs 95% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where DeepSeek-R1 pulls ahead
- Reasons across sets of long documentsAA-LCR+54.4points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+37points ahead
- Code for real scientific research problemsSciCode+28.2points ahead
Where Granite 3.3 8B Instruct pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 28 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-R1
- SciCodeDeepSeek-R1 by 28.238.310.1+28.2
- Terminal-Bench HardDeepSeek-R1 by 6.16.10+6.1
ReasoningDeepSeek-R1
- GPQA DiamondDeepSeek-R1 by 3770.833.8+37
- Humanity's Last ExamDeepSeek-R1 by 4.38.54.2+4.3
- CritPttie0.60tie
FactsDeepSeek-R1
- AA-Omniscience · AccuracyDeepSeek-R1 by 20.830.59.7+20.8
- AA-Omniscience · Non-hallucinationDeepSeek-R1 by 5.99.73.8+5.9
- Vectara HHEM hallucination ratelower is bettertie11.310.6tie
Following instructionsDeepSeek-R1
- IFBenchDeepSeek-R1 by 16.63922.4+16.6
Other results17 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 2024DeepSeek-R1 by 71.779.88.1+71.7
- AA-OmniscienceDeepSeek-R1 by 45.1-32.2-77.3+45.1
- HarmBenchGranite 3.3 8B Instruct by 33.347.981.2+33.3
- AIR-Bench 2024Granite 3.3 8B Instruct by 29.652.982.5+29.6
- MATH-500 (EM)DeepSeek-R1 by 28.397.369+28.3
- MMLUDeepSeek-R1 by 25.390.865.5+25.3
- Artificial Analysis Coding IndexDeepSeek-R1 by 21.224.63.4+21.2
- bbqDeepSeek-R1 by 12.696.684+12.6
- IFEvalDeepSeek-R1 by 8.583.374.8+8.5
- AA IntelligenceDeepSeek-R1 by 6.511.44.9+6.5
- vectara_avg_summary_lengthGranite 3.3 8B Instruct by 37.9 rating points93.5131.4+37.9 rating
- XSTestDeepSeek-R1 by 3.794.490.7+3.7
- vectara_answer_rateGranite 3.3 8B Instruct by 397100+3
- anthropic_red_teamGranite 3.3 8B Instruct by 1.997.299.1+1.9
- τ²-Bench Telecom (AA run)tie11.410.5tie
- vectara_factual_consistencytie88.789.4tie
- simple_safety_teststie9897.5tie
Questions people ask
Which is better, DeepSeek-R1 or Granite 3.3 8B Instruct?
DeepSeek-R1 wins five of the six areas where both have results: coding, reasoning, facts, long documents and following instructions. Granite 3.3 8B Instruct wins none, but costs 95% less. They are level on agents.
Which is better for coding?
DeepSeek-R1. It wins 2 of the 2 coding tests both models report; Granite 3.3 8B Instruct wins none.
Which is cheaper?
DeepSeek-R1 costs $2.00 per million input tokens and $4.00 per million output tokens; Granite 3.3 8B Instruct costs $0.03 and $0.25. That makes Granite 3.3 8B Instruct about 95% cheaper for the same work.
How do you compare the two?
We use the 28 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 17 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.