Head to head
Phi-4 vs Gemma 3 12B
Phi-4 from Microsoft against Gemma 3 12B from Google DeepMind — specification, price and every benchmark both makers have published, in one table.
Benchmarks
Phi-4 leads
Phi-4 wins 2 of the 2 benchmarks both models report, Gemma 3 12B wins 0, by a wide margin. The average gap across shared tests is 15.5 points.
Price
Gemma 3 12B is cheaper
On a 3:1 input-to-output mix, Gemma 3 12B costs $0.062 per million tokens against $0.088 for Phi-4 — about 1.4× less.
What actually differs
- Gemma 3 12B takes 131K tokens of context against 16K — 8.0× more room for long documents or a large codebase.
- Only Gemma 3 12B reads images. If your input includes screenshots, charts or documents, that decides it.
Scorecard
Which is better at what
Maths, coding, reasoning and the rest — one line each, averaged over the benchmarks both models actually report.
| Category | Phi-4 | Gemma 3 12B | Better at this |
|---|---|---|---|
| ReasoningGPQA Diamond | 56.1% | 34.9% | Phi-4+21.2 |
| MathsNeither model reports this | — | — | Not comparable |
| CodingOnly one model reports this | — | — | Not comparable |
| KnowledgeMMLU-Pro | 70.4% | 60.6% | Phi-4+9.8 |
| MultimodalOnly one model reports this | — | — | Not comparable |
| Instruction followingNeither model reports this | — | — | Not comparable |
| Human preferenceNeither model reports this | — | — | Not comparable |
| Categories wonOut of 2 comparable | 2 | 0 | Phi-4 |
Each category averages only the benchmarks every model here reports, so no one is credited for a test the other did not run. A category with no shared test is marked Not comparable rather than guessed at.
Side by side
Phi-4 and Gemma 3 12B, row by row
| Attribute | Phi-4Microsoft | Gemma 3 12BGoogle DeepMind |
|---|---|---|
| Specification | ||
| MakerWho built it | Microsoft | Google DeepMind |
| Released | 2024-12 | 2025-03 |
| ParametersTotal, and active per token for a mixture of experts | 14B | 12B |
| Architecture | Dense transformer | Dense transformer |
| Context windowHow much can go in at once | 16,384 tokens | 131,072 tokens |
| Max output | 4,096 tokens | 8,192 tokens |
| Input | Text | Text, Image |
| ReasoningSpends extra tokens thinking before it answers | No | No |
| Tool calling | No | Yes |
| Licence | MIT | Gemma Terms of Use |
| Open weightsCan you download and run it yourself | Yes | Yes |
| Price | ||
| Input priceUSD per million tokens in | $0.070 | $0.05 |
| Output priceUSD per million tokens out | $0.140 | $0.10 |
| Blended 3:1A 3-in-to-1-out million-token mix — a fairer single number than input price alone | $0.088 | $0.062Cheapest |
| Price note | Open weights — this is a representative hosting rate, not a first-party price. Running it yourself costs only hardware. | Open weights — this is a representative hosting rate, not a first-party price. Running it yourself costs only hardware. |
| Published benchmarks | ||
| MMLU-Pro12,000 reasoning-heavy multiple-choice questions across 14 academic subjects, with ten options instead of four. The harder successor to MMLU. | ||
| GPQA Diamond198 graduate-level physics, chemistry and biology questions written to be Google-proof. PhD holders in the matching field score about 65%. | ||
| HumanEval164 short Python functions written from a docstring. Saturated at the frontier — kept here for continuity with older models. | Not reported | |
| MMMUCollege-level questions that require reading charts, diagrams, tables and photographs alongside the text. | Not reported | |
| Links | Hugging Face · Full page | Hugging Face · Full page |
| Row verified | 2026-08 | 2026-08 |
Where these numbers come from
Every score on this page is a published figure, taken from the model's own card, system card, technical report or release post, or from a public leaderboard. CorX Labs did not run these evaluations. Most are self-reported by the lab that built the model, which means they were produced under that lab's own choice of prompt, scaffold and number of attempts — so treat them as a starting point for a shortlist, not as a settled ranking.
A score someone other than the model's maker measured is marked Independent and names its measurer. Those are the stronger numbers on this page — an outside harness has no reason to flatter anyone — and there are not many of them.
Where a figure has not been published, the cell reads Not reported rather than an estimate. Nothing here is inferred, interpolated or guessed. Each model records the month its row was last checked. Full method and caveats.
Questions
Phi-4 or Gemma 3 12B?
Is Phi-4 better than Gemma 3 12B?
Phi-4 wins 2 of the 2 benchmarks both models report, Gemma 3 12B wins 0, by a wide margin. The average gap across shared tests is 15.5 points.
Which is cheaper, Phi-4 or Gemma 3 12B?
On a 3:1 input-to-output mix, Gemma 3 12B costs $0.062 per million tokens against $0.088 for Phi-4 — about 1.4× less.
What is the difference between Phi-4 and Gemma 3 12B?
Gemma 3 12B takes 131K tokens of context against 16K — 8.0× more room for long documents or a large codebase. Only Gemma 3 12B reads images. If your input includes screenshots, charts or documents, that decides it.
Related
