Automated page. Written and translated automatically by AI and checked against the sources by software; no human editor. SI-IQ is an index built from public benchmark results, not an IQ test. How it works
How smart is Phi-4?
Phi-4 is a model from Microsoft Research, released on Dec 12, 2024. Its SI-IQ is 76 (90% interval 71–81), estimated from 4 benchmark results by 2 evaluators.
Data as of Oct 18, 2026 · method 1.0.0
A model is ranked when its results come from at least 5 benchmarks (the two GPQA Diamond runs count as one) in at least 3 areas, from at least 2 different evaluators, and its versions are released no more than 60 days apart. This model does not meet that yet, so it is shown with a provisional estimate but not ranked.
Facts
- Developer
- Microsoft Research
- Released
- Dec 12, 2024
- Access
- Open weights (restricted licence)
- Benchmarks used
- 4 benchmark results from 2 evaluators
Ability by area (SI-IQ, 90% interval)
Every benchmark result we used
Each score is one value exactly as the source published it: for Epoch AI the best of its run settings, for OpenRouter the median hosting provider (the developer's own hosting when it ran the test). “Implied SI-IQ” is the ability that this single result points to; the overall SI-IQ is their precision-weighted average.
| Benchmark | Score | Evaluated by | Source |
|---|---|---|---|
| MATH Level 5 | 64.9% | Epoch AI | Epoch AI Benchmarking Hub · data file |
| OTIS Mock AIME 2024–2025 | 13.8% | Epoch AI | Epoch AI Benchmarking Hub · data file |
| GPQA Diamond | 56.1% | Epoch AI | Epoch AI Benchmarking Hub · data file |
| Berkeley Function Calling Leaderboard | 28.8% | UC Berkeley (Gorilla) | Berkeley Function Calling Leaderboard · data file |
Human preference (Arena)
On the Arena leaderboard, where people vote between two anonymous answers, Phi-4 has a rating of 1,256 (1,252–1,261) from 24,126 votes (published Oct 15, 2026). Human preference is a different measure from benchmark ability and is not part of SI-IQ. Arena (Hugging Face)
Computed with SI-IQ method version 1.0.0. Read the full methodology
Data sources: Arena leaderboard dataset on Hugging Face (CC BY 4.0); Berkeley Function Calling Leaderboard, Gorilla project (Apache-2.0); Epoch AI, ‘Capabilities & benchmarking’ (CC BY 4.0) (licences and changes).