vs
0:2
overall capability wins (2 ties across 4 scored). Gemma 3 27B leads overall.
- Arena0:1
- Pricing1:2
- Speed0:2
- Specs0:1
Verdict
Gemma 3 27B ranks higher on LMArena (+76 Elo); Gemma 3 27B offers ~15.6x larger context (128K vs 8K); Gemma 3 27B is ~1.8x cheaper on a blended token basis than Gemma 2 27B.
- Gemma 3 27B ranks higher on LMArena (+76 Elo)
- Gemma 3 27B offers ~15.6x larger context (128K vs 8K)
- Gemma 3 27B is ~1.8x cheaper on a blended token basis than Gemma 2 27B
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Gemma 2 27B
- $0.34
- Gemma 3 27B
- $0.17
Gemma 3 27B is about 1.9x 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.)
Google AI / self-host
Modalities
text → text
text, image → text
Specs
Context window
8K
128K
Max output
8K
8K
Parameters
27B
27B
Pricing
Input $/1M
$0.27
$0.10
Output $/1M
$0.27
$0.30
Blended $/1M (3∶1)
$0.27
$0.15
Speed
tok/s
80
90
TTFT (s)
0.35
0.28
Reasoning
MMLU-Pro
—
67.5
GPQA Diamond
—
42.8
Humanity's Last Exam
—
4.4
AIME 2025
—
20.7
MATH-500
—
88.3
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
13.0
AA-LCR v1.1
—
7.3
CritPt
—
—
MMMU-Pro
—
48.0
IFBench
—
31.8
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
11.4
SWE-bench Multilingual
—
—
LiveCodeBench
—
29.7
Terminal-Bench 2.1
—
4.5
Aider Polyglot
—
4.9
Terminal-Bench 3
—
—
BigCodeBench
42.8
—
SciCode
—
23.3
CursorBench
—
—
SWE-Rebench
—
5.4
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,289
1,365
| Metric | Gemma 2 27B | Gemma 3 27B | 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.) | Google AI / self-host | — |
| Modalities | text → text | text, image → text | — |
| Specs | |||
| Context window | 8K | 128K | B +120K |
| Max output | 8K | 8K | tie |
| Parameters | 27B | 27B | tie |
| Pricing | |||
| Input $/1M | $0.27 | $0.10 | B +$0.17 |
| Output $/1M | $0.27 | $0.30 | A +$0.03 |
| Blended $/1M (3∶1) | $0.27 | $0.15 | B +$0.12 |
| Speed | |||
| tok/s | 80 | 90 | B +10 |
| TTFT (s) | 0.35 | 0.28 | B +0.07 |
| Reasoning | |||
| MMLU-Pro | — | 67.5 | — |
| GPQA Diamond | — | 42.8 | — |
| Humanity's Last Exam | — | 4.4 | — |
| AIME 2025 | — | 20.7 | — |
| MATH-500 | — | 88.3 | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 13.0 | — |
| AA-LCR v1.1 | — | 7.3 | — |
| CritPt | — | — | — |
| MMMU-Pro | — | 48.0 | — |
| IFBench | — | 31.8 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | 11.4 | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 29.7 | — |
| Terminal-Bench 2.1 | — | 4.5 | — |
| Aider Polyglot | — | 4.9 | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 42.8 | — | — |
| SciCode | — | 23.3 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | 5.4 | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,289 | 1,365 | B +76 Elo |
Gemma 2 27B
- Context
- 8K / 8K out
- Parameters
- 27B
- Price
- $0.27 / $0.27
- Speed
- 80 tok/s · 0.35s TTFT
- Modalities
- text → text
- License
- Gemma
Google's open 27B instruct model — strong small-model quality before Gemma 3 / 4.
Gemma 3 27B
- Context
- 128K / 8K out
- Parameters
- 27B
- Price
- $0.10 / $0.30
- Speed
- 90 tok/s · 0.28s TTFT
- Modalities
- text, image → text
- License
- Gemma
Open multimodal Gemma 3 mid-size - strong preference Elo for its parameter count.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.