vs
1:2
overall capability wins (1 ties across 4 scored). Gemma 3 4B leads overall.
- Arena0:1
- Pricing0:2
- Speed0:2
- Specs1:1
Verdict
Gemma 3 4B ranks higher on LMArena (+36 Elo); Gemma 3 4B offers ~15.6x larger context (128K vs 8K); Gemma 3 4B is ~1.6x cheaper on a blended token basis than Gemma 2 9B.
- Gemma 3 4B ranks higher on LMArena (+36 Elo)
- Gemma 3 4B offers ~15.6x larger context (128K vs 8K)
- Gemma 3 4B is ~1.6x cheaper on a blended token basis than Gemma 2 9B
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Gemma 2 9B
- $0.10
- Gemma 3 4B
- $0.06
Gemma 3 4B is about 1.7x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Family
Gemma
Gemma
License
Gemma
Gemma
Open weights
Yes
Yes
Release
Jun 27, 2024
Mar 12, 2025
Knowledge cutoff
-
-
API / provider
Groq / Together (ref.)
Together / Fireworks (ref.)
Modalities
text → text
text, image → text
Specs
Context window
8K
128K
Max output
8K
8K
Parameters
9B
4B
Pricing
Input $/1M
$0.08
$0.04
Output $/1M
$0.08
$0.08
Blended $/1M (3∶1)
$0.08
$0.05
Speed
tok/s
140
160
TTFT (s)
0.2
0.18
Reasoning
MMLU-Pro
—
43.6
GPQA Diamond
—
29.1
Humanity's Last Exam
—
5.3
AIME 2025
—
12.7
MATH-500
—
76.6
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
7.7
AA-LCR v1.1
—
6.7
CritPt
—
—
MMMU-Pro
—
29.9
IFBench
—
28.3
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
12.6
Terminal-Bench 2.1
—
0.4
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
34.7
—
SciCode
—
7.3
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,267
1,303
| Metric | Gemma 2 9B | Gemma 3 4B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | — | ||
| Family | Gemma | Gemma | — |
| License | Gemma | Gemma | — |
| Open weights | Yes | Yes | — |
| Release | Jun 27, 2024 | Mar 12, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Groq / Together (ref.) | Together / Fireworks (ref.) | — |
| Modalities | text → text | text, image → text | — |
| Specs | |||
| Context window | 8K | 128K | B +120K |
| Max output | 8K | 8K | tie |
| Parameters | 9B | 4B | A +5.00 |
| Pricing | |||
| Input $/1M | $0.08 | $0.04 | B +$0.04 |
| Output $/1M | $0.08 | $0.08 | tie |
| Blended $/1M (3∶1) | $0.08 | $0.05 | B +$0.03 |
| Speed | |||
| tok/s | 140 | 160 | B +20 |
| TTFT (s) | 0.2 | 0.18 | B +0.02 |
| Reasoning | |||
| MMLU-Pro | — | 43.6 | — |
| GPQA Diamond | — | 29.1 | — |
| Humanity's Last Exam | — | 5.3 | — |
| AIME 2025 | — | 12.7 | — |
| MATH-500 | — | 76.6 | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 7.7 | — |
| AA-LCR v1.1 | — | 6.7 | — |
| CritPt | — | — | — |
| MMMU-Pro | — | 29.9 | — |
| IFBench | — | 28.3 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 12.6 | — |
| Terminal-Bench 2.1 | — | 0.4 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 34.7 | — | — |
| SciCode | — | 7.3 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,267 | 1,303 | B +36 Elo |
Gemma 2 9B
- Context
- 8K / 8K out
- Parameters
- 9B
- Price
- $0.08 / $0.08
- Speed
- 140 tok/s · 0.2s TTFT
- Modalities
- text → text
- License
- Gemma
Efficient open 9B Gemma 2 instruct — popular local and edge chat baseline.
Gemma 3 4B
- Context
- 128K / 8K out
- Parameters
- 4B
- Price
- $0.04 / $0.08
- Speed
- 160 tok/s · 0.18s TTFT
- Modalities
- text, image → text
- License
- Gemma
Small multimodal Gemma 3 with long context — strong on-device and edge deployment option.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.