Gemma 4 120B
Gemma 4 120B는 Google DeepMind가 개발한 기초 모델입니다. 128K의 컨텍스트 길이를 특징으로 하며, 대규모 파라미터 규모로 인한 고급 처리 능력이 특징입니다.
파라미터
1200.0B
컨텍스트
128K
라이선스
Gemma License
출시일
2026-05-19
API 가격
이 모델의 API 가격 정보는 현재 공개되지 않았습니다
강점
- ・대규모 1200억 파라미터
- ・128K의 긴 컨텍스트 윈도우
- ・Google DeepMind 개발
약점
- ・모델 규모로 인한 높은 계산 비용
- ・증가한 메모리 소비
- ・추론 속도에 대한 영향
활용 사례
- ・대규모 데이터 분석
- ・장문 문서 처리
- ・강력한 기초 모델로의 활용
심층 분석
Note
No 120B variant exists
Gemma 4 comes in E2B (2.3B eff), E4B (4.5B eff), 26B MoE, and 31B Dense - NOT 120B
Largest Model
31B Dense
The largest Gemma 4 variant is 31B dense parameters
Arena AI Text Ranking
#3 open model (31B)
Estimated LMArena score 1452 for 31B model
26B MoE Ranking
#6 open model
Arena score 1441 with only 4B active parameters
Context Window
256K (31B/26B) / 128K (E2B/E4B)
Long context for the larger models
Release Date
April 2, 2026
Announced by Google DeepMind
License
Apache 2.0
First Gemma model with Apache 2.0 license
Languages
140+
Natively trained on over 140 languages
VRAM (31B bfloat16)
Single 80GB H100
Fits on one H100; quantized versions run on consumer GPUs
강점
- ・Apache 2.0 license - first Gemma model with truly permissive licensing
- ・31B model ranks #3 open model on Arena AI text leaderboard, outcompeting models 20x its size
- ・26B MoE activates only 4B parameters during inference - massive efficiency
- ・Multimodal: vision and audio input on all variants, including edge models
- ・140+ language support natively trained
- ・Strong agentic capabilities: function calling, structured JSON, system instructions
약점
- ・No 120B variant exists - largest is 31B (requester's slug is incorrect)
- ・31B dense requires 80GB H100 for bfloat16 (quantized needed for consumer GPUs)
- ・E2B is text-only with 8K context (limited compared to larger variants)
- ・No thinking mode on E2B variant
- ・Relatively new (April 2026), limited production deployment data
경쟁사 비교
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| Gemma 4 31B Dense | 1452 (est) | ~65% (est) | ~75% (est) | Free (open weights) |
| Gemma 4 26B MoE (4B active) | 1441 (est) | ~60% (est) | ~70% (est) | Free (open weights) |
| Llama 4 Maverick (400B MoE) | ~1460 | ~70% | ~80% | Free (open weights) |
| Qwen3 235B MoE | ~1450 | ~68% | ~78% | Free (open weights) |
| DeepSeek-V3.5 | ~1455 | ~72% | ~82% | Free (open weights) |
Gemma 4 is Google's most capable open model family (April 2, 2026), built from Gemini 3 research. Available in E2B, E4B, 26B MoE, and 31B Dense sizes under Apache 2.0 license. The 31B model ranks #3 among open models on Arena AI, outcompeting models 20x its size. Note: there is no 120B variant - the largest model is 31B dense.
출처
분석 생성일: 2026-05-24