Menu

Best local models for Hermes Agent by memory

Pick your memory size and see which models get Hermes Agent tasks done, ranked by measured task completion.

Best for 16 GB

RTX 4060 Ti 16GB / 4070 Ti Super / 5070 Ti, 16–24 GB Macs

  1. gpt-oss-20b MXFP4 (official GGUF) 90%
  2. Qwopus-GLM-18B Q4_K_M (GGUF, healed) 90%

Best for 24 GB

RTX 3090 / 4090, 32 GB RTX 5090, 32–48 GB Macs

  1. Nemotron-Cascade-2-30B-A3B Q4_K_M (GGUF) 100%
  2. GLM-4.7-Flash Q4_K_M (GGUF) 93.3%
  3. Hermes-4-14B Q4_K_M (GGUF) 40%
  4. Gemma 4 31B NVFP4 Not tested yet
  5. Qwen3.8 35B-A3B NVFP4 Not tested yet

Best for 128 GB

DGX Spark and other GB10 machines, Strix Halo 128 GB, 96–128 GB Macs

  1. Gemma 4 26B-A4B Q8_0 (GGUF) 100%
  2. gpt-oss-120b MXFP4 (official GGUF) 100%
  3. Qwen3.6-35B-A3B UD-Q6_K (GGUF) 100%
  4. Qwen3.6-27B Q6_K (GGUF) 96.7%
  5. Nemotron-3-Nano-30B-A3B Q8_0 (GGUF) 83.3%
  6. Hermes-4-70B Q4_K_M (GGUF) 70%
  7. Hermes-4.3-36B Q6_K (GGUF) 70%
  8. Mistral Small 4 119B NVFP4 Not tested yet
  9. Nemotron-3-Super-120B-A12B NVFP4 Not tested yet
  10. Qwen3-Coder-Next FP8 Not tested yet
  11. Qwen3.5-122B-A10B NVFP4 Not tested yet

Measured on a 128 GB DGX Spark. Smaller tiers are grouped by expected memory use (weights plus context). Measured server memory replaces the estimate once a setup is tested. Speeds on your own GPU will differ.