Qwen 2.5 72B Instruct
Qwen · Qwen License
Top-tier 72B model, competitive with Llama 3.1 70B. Capable at coding and structured output. Requires 48 GB+ VRAM for usable quantizations.
- Parameters
- 72.7B
- Architecture
- Dense
- Context
- 131,072 tokens
- Released
- 2024-09-19
- Engines
- llama.cpp, ollama, vLLM
Parameters
72.7B
VRAM
40.5 GB
Context
128K
Formats
3
GPUs
19
Qwen 2.5 72B Instruct (72.7B) requires 40.5 GB VRAM at recommended quality (Q4_K_M). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 60 tok/s at Q4_K_M. For the best experience, High-End Home AI Server ($3,842) is recommended.
Source: OwnRig methodology
40.5 GB
Q4_K_M
36.4 GB
128K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| efficient | Q4_K_M | 40.5 GB | 36.4 GB |
| compressed | Q3_K_M | 32.5 GB | 28.4 GB |
| compressed | Q2_K | 25.3 GB | 21.8 GB |
Context Length Impact
KV cache VRAM at Q4_K_M quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 717 MB | 41.2 GBexceeds 24 GB |
| 4K | 1.3 GB | 41.8 GBexceeds 24 GB |
| 8K | 2.6 GB | 43.1 GBexceeds 24 GB |
| 16K | 5.3 GB | 45.8 GBexceeds 24 GB |
| 32K | 10.6 GB | 51.1 GBexceeds 24 GB |
| 64K | 21.1 GB | 61.6 GBexceeds 24 GB |
| 128K | 42.2 GB | 82.7 GBexceeds 24 GB |
Compatible GPUs
19 devices| NVIDIA Grace Blackwell Ultra GB300 | Q4_K_M | 60 tok/s | Excellent |
| Apple M4 Max (128GB Unified) | Q4_K_M | 6 tok/s | Acceptable |
| Apple M4 Max (64GB Unified) | Q3_K_M | 6 tok/s | Acceptable |
| Apple M4 Ultra (192GB) | Q4_K_M | 9 tok/s | Acceptable |
| AMD Radeon Pro W7900 | Q3_K_M | 6 tok/s | Acceptable |
| NVIDIA RTX PRO 6000 Blackwell | Q4_K_M | 26 tok/s | Acceptable |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Q4_K_M | 24 tok/s | Acceptable |
| AMD Radeon RX 7900 XTX | Q2_K | 2 tok/s | Marginal |
| Apple M3 Pro (18GB Unified) | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 3080 10GB | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 4060 8GB | Q2_K | – | Not viable |
| NVIDIA RTX 4060 Laptop (40-60W) | Q2_K | – | Not viable |
| NVIDIA RTX 4070 Laptop (80-115W) | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 4070 Ti 12GB | Q2_K | – | Not viable |
| AMD Radeon RX 7600 | Q2_K | – | Not viable |
| AMD Radeon RX 9070 | Q2_K | – | Not viable |
| AMD Radeon RX 9060 XT 16GB | Q2_K | – | Not viable |
| AMD Radeon RX 9060 XT 8GB | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 5060 8GB | Q2_K | – | Not viable |
Showing 19 of 19 entries
Recommended Builds
Complete PC builds that can run Qwen 2.5 72B Instruct.
Extreme AI Workstation
Dual GPUs that run the biggest AI models at a smart price
Runs 8 models
High-End Home AI Server
Your household's private AI: chatbots, code tools, and more
Runs 12 models
Frequently Asked Questions
- How much VRAM does Qwen 2.5 72B Instruct need?
- Qwen 2.5 72B Instruct requires 40.5 GB VRAM at recommended quality (Q4_K_M). At lower quality settings, it can fit in as little as 25.3 GB.
- What is the best GPU for Qwen 2.5 72B Instruct?
- The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for Qwen 2.5 72B Instruct, achieving 60 tok/s at Q4_K_M with an excellent rating.
- What quantization should I use for Qwen 2.5 72B Instruct?
- For the best quality, use Q4_K_M (40.5 GB VRAM). If your GPU has limited VRAM, Q2_K (25.3 GB) is the most efficient option with acceptable quality.
- Is Qwen 2.5 72B Instruct good for coding?
- Qwen 2.5 72B Instruct supports coding use cases. For the best coding experience, pair it with an embedding model for local RAG.
Related Guides
Explainer
VRAM: The Only Spec That Matters for AI
VRAM for local AI: what it is, why models need it, how quantization cuts requirements, and a VRAM table for major models.
Explainer
How we test: OwnRig's benchmark methodology
How OwnRig measures tokens per second, rates model compatibility, and keeps hardware data current. Our methodology, tools, and known limitations.
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.Qwen is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.