vs
0:3
overall capability wins (1 ties across 4 scored). Gemma 3 12B leads overall.
- Arena0:1
- Pricing3:0
- Speed2:0
- Specs0:2
Verdict
Gemma 3 12B ranks higher on LMArena (+142 Elo); Gemma 3 12B offers ~15.6x larger context (128K vs 8K); Gemma 2 2B is ~1.9x cheaper on a blended token basis than Gemma 3 12B.
- Gemma 3 12B ranks higher on LMArena (+142 Elo)
- Gemma 3 12B offers ~15.6x larger context (128K vs 8K)
- Gemma 2 2B is ~1.9x cheaper on a blended token basis than Gemma 3 12B
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Gemma 2 2B
- $0.05
- Gemma 3 12B
- $0.09
Gemma 2 2B 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
Jul 31, 2024
Mar 12, 2025
Knowledge cutoff
-
-
API / provider
Groq / Together (ref.)
Google AI / self-host
Modalities
text → text
text, image → text
Specs
Context window
8K
128K
Max output
8K
8K
Parameters
2B
12B
Pricing
Input $/1M
$0.04
$0.05
Output $/1M
$0.04
$0.15
Blended $/1M (3∶1)
$0.04
$0.08
Speed
tok/s
200
120
TTFT (s)
0.12
0.22
Reasoning
MMLU-Pro
—
60.6
GPQA Diamond
—
34.9
Humanity's Last Exam
—
4.2
AIME 2025
—
18.3
MATH-500
—
85.3
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
10.3
AA-LCR v1.1
—
8.3
CritPt
—
—
MMMU-Pro
—
37.5
IFBench
—
36.7
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
24.6
Terminal-Bench 2.1
—
0.0
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
—
16.4
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,200
1,342
| Metric | Gemma 2 2B | Gemma 3 12B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | — | ||
| Family | Gemma | Gemma | — |
| License | Gemma | Gemma | — |
| Open weights | Yes | Yes | — |
| Release | Jul 31, 2024 | Mar 12, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Groq / Together (ref.) | Google AI / self-host | — |
| Modalities | text → text | text, image → text | — |
| Specs | |||
| Context window | 8K | 128K | B +120K |
| Max output | 8K | 8K | tie |
| Parameters | 2B | 12B | B +10 |
| Pricing | |||
| Input $/1M | $0.04 | $0.05 | A +$0.01 |
| Output $/1M | $0.04 | $0.15 | A +$0.11 |
| Blended $/1M (3∶1) | $0.04 | $0.08 | A +$0.04 |
| Speed | |||
| tok/s | 200 | 120 | A +80 |
| TTFT (s) | 0.12 | 0.22 | A +0.10 |
| Reasoning | |||
| MMLU-Pro | — | 60.6 | — |
| GPQA Diamond | — | 34.9 | — |
| Humanity's Last Exam | — | 4.2 | — |
| AIME 2025 | — | 18.3 | — |
| MATH-500 | — | 85.3 | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 10.3 | — |
| AA-LCR v1.1 | — | 8.3 | — |
| CritPt | — | — | — |
| MMMU-Pro | — | 37.5 | — |
| IFBench | — | 36.7 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 24.6 | — |
| Terminal-Bench 2.1 | — | 0.0 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | — | 16.4 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,200 | 1,342 | B +142 Elo |
Gemma 2 2B
- Context
- 8K / 8K out
- Parameters
- 2B
- Price
- $0.04 / $0.04
- Speed
- 200 tok/s · 0.12s TTFT
- Modalities
- text → text
- License
- Gemma
Tiny open Gemma 2 instruct — ultra-cheap edge and high-throughput classification baseline.
Gemma 3 12B
- Context
- 128K / 8K out
- Parameters
- 12B
- Price
- $0.05 / $0.15
- Speed
- 120 tok/s · 0.22s TTFT
- Modalities
- text, image → text
- License
- Gemma
Compact open Gemma 3 instruct model for edge and low-cost multimodal chat.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.