Gemma 2 27B Instruct
Gemma · Gemma License
Google's 27B model with effective knowledge distillation. Reasoning and coding at a size that fits on a single 24 GB GPU at Q4. Limited to 8K context.
- Parameters
- 27.23B
- Architecture
- Dense
- Context
- 8,192 tokens
- Released
- 2024-06-27
- Engines
- llama.cpp, ollama, vLLM
Parameters
27.23B
VRAM
18.5 GB
Context
8K
Formats
4
GPUs
21
Gemma 2 27B Instruct (27.23B) requires 18.5 GB VRAM at recommended quality (Q5_K_M). At efficient quality (Q4_K_M), it fits in 15.5 GB VRAM, making it compatible with the NVIDIA RTX 4090 Laptop (150-175W). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 145 tok/s at Q5_K_M. For the best experience, AMD AI Powerhouse ($1,818) is recommended.
Source: OwnRig methodology
18.5 GB
Q5_K_M
16.3 GB
8K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| recommended | Q5_K_M | 18.5 GB | 16.3 GB |
| efficient | Q4_K_M | 15.5 GB | 13.6 GB |
| compressed | Q3_K_M | 12.5 GB | 10.6 GB |
| compressed | Q2_K | 9.8 GB | 8.2 GB |
Context Length Impact
KV cache VRAM at Q5_K_M quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 307 MB | 18.8 GB |
| 4K | 614 MB | 19.1 GB |
| 8K | 1.3 GB | 19.8 GB |
Compatible GPUs
21 devices| NVIDIA Grace Blackwell Ultra GB300 | Q5_K_M | 145 tok/s | Excellent |
| Apple M4 Max (36GB Unified) | Q5_K_M | 15 tok/s | Good |
| NVIDIA GeForce RTX 4090 | Q4_K_M | 22 tok/s | Good |
| AMD Radeon RX 7900 XTX | Q4_K_M | 19 tok/s | Good |
| AMD Radeon Pro W7900 | Q4_K_M | 24 tok/s | Good |
| NVIDIA RTX PRO 6000 Blackwell | Q5_K_M | 59 tok/s | Good |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Q5_K_M | 54 tok/s | Good |
| NVIDIA GeForce RTX 4060 Ti 16GB | Q4_K_M | 12 tok/s | Acceptable |
| NVIDIA RTX 4090 Laptop (150-175W) | Q4_K_M | 10 tok/s | Acceptable |
| AMD Radeon RX 9070 | Q3_K_M | 22 tok/s | Acceptable |
| AMD Radeon RX 9060 XT 16GB | Q3_K_M | 11 tok/s | Acceptable |
| NVIDIA GeForce RTX 5060 Ti 16GB | Q4_K_M | 13 tok/s | Acceptable |
| AMD Radeon RX 7600 | Q2_K | 2 tok/s | Marginal |
| Apple M3 Pro (18GB Unified) | Q4_K_M | – | Not viable |
| NVIDIA GeForce RTX 3080 10GB | Q3_K_M | – | Not viable |
| NVIDIA GeForce RTX 4060 8GB | Q3_K_M | – | Not viable |
| NVIDIA RTX 4060 Laptop (40-60W) | Q3_K_M | – | Not viable |
| NVIDIA RTX 4070 Laptop (80-115W) | Q3_K_M | – | Not viable |
| NVIDIA GeForce RTX 4070 Ti 12GB | Q3_K_M | – | Not viable |
| AMD Radeon RX 9060 XT 8GB | Q3_K_M | – | Not viable |
| NVIDIA GeForce RTX 5060 8GB | Q3_K_M | – | Not viable |
Showing 21 of 21 entries
Recommended Builds
Complete PC builds that can run Gemma 2 27B Instruct.
AI Builder Workstation
Run every AI tool you need. Nothing leaves your machine
Runs 10 models
High-End AI Workstation
Chat, generate images, and code with AI, all at once
Runs 8 models
High-End Home AI Server
Your household's private AI: chatbots, code tools, and more
Runs 12 models
Mid-Range Home AI Server
Serve multiple AI models to every device at home
Runs 9 models
Frequently Asked Questions
- How much VRAM does Gemma 2 27B Instruct need?
- Gemma 2 27B Instruct requires 18.5 GB VRAM at recommended quality (Q5_K_M). At lower quality settings, it can fit in as little as 9.8 GB.
- What is the best GPU for Gemma 2 27B Instruct?
- The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for Gemma 2 27B Instruct, achieving 145 tok/s at Q5_K_M with an excellent rating.
- Can I run Gemma 2 27B Instruct on an RTX 4060 Ti?
- Yes. On the NVIDIA GeForce RTX 4060 Ti 16GB, Gemma 2 27B Instruct runs at 12 tok/s (Q4_K_M, acceptable).
- What quantization should I use for Gemma 2 27B Instruct?
- For the best quality, use Q5_K_M (18.5 GB VRAM). If your GPU has limited VRAM, Q2_K (9.8 GB) is the most efficient option with acceptable quality.
- Is Gemma 2 27B Instruct good for coding?
- Gemma 2 27B Instruct supports coding use cases. For the best coding experience, pair it with an embedding model for local RAG.
Data confidence: verified. Source
VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.Gemma is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.