====== LLM ======

- https://collabnix.com/best-ollama-models-in-2025-complete-performance-comparison/

For Production Deployment:
  * Primary Choice: DeepSeek-R1 32B for reasoning-heavy applications
  * Coding Tasks: Qwen2.5-Coder 7B for optimal balance of capability and efficiency
  * General Purpose: Llama 3.3 70B for maximum versatility
  * Edge Computing: Phi-4 14B for resource-constrained environments

Optimization Strategies:
  * Always enable **Flash Attention** and KV-cache quantization
  * Use **Q4_K_M** quantization for production deployments
  * Implement caching for repeated queries
  * Monitor GPU memory usage and implement automatic model swapping
  * Use load balancing for high-throughput applications



^ Hardware ^ Llama 3.3 8B (tokens/sec) ^ Llama 3.3 70B (tokens/sec) ^ Llama 3.2 ^
| RTX 4090 | 89.2 | 12.1 | |
| RTX 3090 | 67.4 | 8.3 | |
| A100 40GB | 156.7 | 45.2 | |
| M3 Max 128GB | 34.8 | 4.2 | |
| Strix Halo 128GB ollama | | 5.1 | 85.02 |
| Strix Halo 128GB llama.cpp | |  | 90 |
| RTX 3060 | | | 131.76 |



^ model                  ^ capabilities             ^ size     ^ context  ^ quantization                                                                      ^ eval rate [token/s]  ^ prompt eval rate [token/s]  ^
| llama3.2               | completion tools         | "3.2B"   | 131072   | "Q4_K_M"                                                                          | 88.14                | 715.43                      |
| ministral-3:14b        | completion vision tools  | "13.9B"  | 262144   | "Q4_K_M"                                                                          | 23.78                | 302.07                      |
| qwen3-coder:30b        | completion tools         | "30.5B"  | 262144   | "Q4_K_M"                                                                          | 73.75                | 72.41                       |
| llama3:70b | completion   | "70.6B"  | 8192     | "Q4" | 5.55 | 9.72 | 
| llava  | completion vision | "7B"   | 32768    | "Q4"  | 49.92                | 207.27                      |
| deepseek-coder-v2:16b  | completion insert        | "15.7B"  | 163840   | "Q4"                                                                            | 84.44                | 111.71                      |
| bjoernb/qwen3-coder-30b-1m:latest  | completion tools | "30.5B"   | 1048576    | "Q4_K_M" | 74.23 | 94.84 |
| freehuntx/qwen3-coder:8b  | completion tools | "8.2B"   | 40960    | "Q4_K_M" | 37.97 | 565.68 |
| networkjohnny/deepseek-coder-v2-lite-base-q4_k_m-gguf:latest  | completion tools | "3.2B"   | 131072    | "Q4_K_M" | 86.02 | 1124.53 |
| phi4-mini  | completion tools | "3.8B"   | 131072    | "Q4_K_M" | 72.24 | 31.37 |
| qwen2.5:7b  | completion tools | "7.6B"   | 32768    | "Q4_K_M" | 42.98 | 153.34 |
| llama3.3:70b-instruct-q4_K_M  | completion tools | "70.6B"   | 131072    | "Q4_K_M" | 5.06 | 15.50 |
| functiongemma  | completion tools | "268.10M" | 32768 | "Q8" | 364.21 | 240.50 |
| danielsheep/Qwen3-Coder-30B-A3B-Instruct-1M-Unsloth  | completion tools | "30.5B"   | 1048576    | "Q4_K_M" | 71.60 | 33.14 |
| gpt-oss:20b  | completion tools thinking | "20.9B"   | 131072    | "MXFP4" | 47.32 | 402.47 |