Llama 3.3 70B Instruct vs Phi 4
Wins 2 of 5 areas
Long documents · Following instructions
Wins 1 of 5 areas
Reasoning
Llama 3.3 70B Instruct wins more areas, narrowly.Phi 4 is cheaper and better at reasoning.
Scores updated · 25 tests both models report · How we compare
Where each one wins
Tests won in each of the five 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.
- Long documentsFinding answers in very long texts10Llama 3.3 70B Instruct1 of 1 test
- Following instructionsDoing exactly what it is asked10Llama 3.3 70B Instruct1 of 1 test
- ReasoningHard problems that need careful thinking01Phi 41 of 3 tests · 2 ties
- CodingWriting and fixing software00Even0 each · 2 ties
- FactsGetting facts right instead of making them up11Even1 each · 1 tie
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.
Phi 4 costs 56% 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 Llama 3.3 70B Instruct pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+23.6points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+4.9points ahead
Where Phi 4 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+9points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+7.7points ahead
Every test, side by side
All 25 tests both models report. The winning score is in its model's colour; marks a score checked independently.
ReasoningPhi 4
- GPQA DiamondPhi 4 by 7.749.857.5+7.7
- Humanity's Last Examtie3.63.8tie
- CritPttie00tie
FactsEven
- AA-Omniscience · Non-hallucinationPhi 4 by 99.818.8+9
- AA-Omniscience · AccuracyLlama 3.3 70B Instruct by 4.918.914.1+4.9
- Vectara HHEM hallucination ratelower is bettertie4.13.7tie
Long documentsLlama 3.3 70B Instruct
- AA-LCRLlama 3.3 70B Instruct by 15.715.70+15.7
Following instructionsLlama 3.3 70B Instruct
- IFBenchLlama 3.3 70B Instruct by 23.647.123.5+23.6
Other results15 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.
- IFEvalLlama 3.3 70B Instruct by 29.192.163+29.1
- τ²-Bench Telecom (AA run)Llama 3.3 70B Instruct by 26.626.60+26.6
- vectara_answer_rateLlama 3.3 70B Instruct by 18.899.580.7+18.8
- SimpleQALlama 3.3 70B Instruct by 17.920.93+17.9
- DROPLlama 3.3 70B Instruct by 14.790.275.5+14.7
- MGSMLlama 3.3 70B Instruct by 10.591.180.6+10.5
- HumanEvalLlama 3.3 70B Instruct by 5.888.482.6+5.8
- vectara_avg_summary_lengthPhi 4 by 56.3 rating points64.6120.9+56.3 rating
- MATHPhi 4 by 3.47780.4+3.4
- AA IntelligenceLlama 3.3 70B Instruct by 1.87.75.9+1.8
- AA-OmniscienceLlama 3.3 70B Instruct by 1.5-54.2-55.7+1.5
- MMLU-ProPhi 4 by 1.568.970.4+1.5
- MMLULlama 3.3 70B Instruct by 1.28684.8+1.2
- Artificial Analysis Coding Indextie11.911.2tie
- vectara_factual_consistencytie95.996.3tie
Questions people ask
Which is better, Llama 3.3 70B Instruct or Phi 4?
Llama 3.3 70B Instruct wins two of the five areas where both have results: long documents and following instructions. Phi 4 wins reasoning, and costs 56% less. They are level on coding and facts.
Which is better for coding?
Neither. All 2 coding tests both models report are ties.
Which is cheaper?
Llama 3.3 70B Instruct costs $0.71 per million input tokens and $0.72 per million output tokens; Phi 4 costs $0.13 and $0.50. That makes Phi 4 about 56% cheaper for the same work.
How do you compare the two?
We use the 25 benchmark tests both models have published scores on. The verdict counts the 10 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 15 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.