Gemma 3n E4B Instruct vs Gemma 4 E4B
Different efforts: non-thinking · thinking
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Gemma 3n E4B Instruct4.5
Gemma 4 E4B3.8
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Gemma 3n E4B Instruct—
Gemma 4 E4B0
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Gemma 3n E4B Instruct—
Gemma 4 E4B0.5
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Gemma 3n E4B Instruct1,299
Gemma 4 E4B—
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Gemma 3n E4B Instruct26.2
Gemma 4 E4B51.4
| Benchmark | Gemma 3n E4B Instruct | Gemma 4 E4B |
|---|---|---|
| Humanity's Last Exam | 4.5 | 3.8 |
| Terminal-Bench 4.0 | — | 0 |
| AutomationBench-AA | — | 0.5 |
| Arena Text · creative writing | 1,299 | — |
| MMMU-Pro (vision) | 26.2 | 51.4 |
Specs
| Spec | Gemma 3n E4B Instruct | Gemma 4 E4B |
|---|---|---|
| Input, per 1M | — | $0.02 |
| Output, per 1M | — | $0.10 |
| Cost per task? | — | $0.008 |
| Speed, tok/s | — | 29 |
| Context | 32K | 128K |
| Weights | Open | Open |
| Released | Jun 2025 | Apr 2026 |
| Memory (4-bit) | 3.9 GB | 4 GB |
| Licence | Gemma terms | Apache-2.0 |
| Commercial use | Restricted | Allowed |
| See images | ✓ | ✓ |
| Read PDFs | — | — |
| Hear audio | — | ✓ |
| Watch video | — | — |
| Call tools | — | ✓ |
| Force a tool call | ? | ? |
| Return strict JSON | Partial | Partial |
| Search the web | ? | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
Gemma 3n E4B Instruct → Gemma 4 E4B
Memory (4-bit): 3.9 GB → 4 GB
Gemma 4 E4B → Gemma 3n E4B Instruct