# LocalAgentBoard

> Which local models actually work with Hermes Agent, tested on real hardware. An independent, hands-on leaderboard of open-weight models, quantizations and engines for running Hermes Agent locally. Each setup runs the same 10 offline agent tasks 3 times inside the real Hermes Agent on an NVIDIA DGX Spark (GB10, 128 GB unified memory); a run passes only if an automatic checker finds the correct result on disk.

Unofficial. Not affiliated with or endorsed by Nous Research. Hermes Agent is a trademark of its owner.

Key facts:
- 17 of 22 setups tested (2026-10-22 to 2026-10-24, llama.cpp (commit a37cc46); vLLM 0.32.0 for NVFP4/FP8 setups, Hermes Agent v0.22.4, v0.22.5).
- Score = tasks passed out of 30 (10 tasks x 3 tries). Scores within 2 points are a statistical tie ("Top group").
- "Best overall" = fastest setup (decode tokens/second) in the top group, excluding setups under review.
- Full data for every setup: https://localagentboard.com/llms-full.txt

## Main pages
- [Leaderboard](https://localagentboard.com/): all tested setups with score, tool-call success, speed, memory and current top picks
- [Best for 16 / 24 / 128 GB](https://localagentboard.com/best-for/): picks by available memory
- [Methodology](https://localagentboard.com/methodology/): tasks, metrics, fixed settings, retest and tie rules
- [News and analysis](https://localagentboard.com/news/): results, retests and guides
- [About](https://localagentboard.com/about/): independence, sponsorship and referral disclosure

## News
- [Best local LLM for Hermes Agent: best overall, fastest and best for small machines](https://localagentboard.com/news/best-local-llm-for-hermes-agent-top-picks/): Which local model should you run Hermes Agent on? Measured picks from 16 tested setups on a DGX Spark: best overall, fastest reliable and best for 16–24 GB machines, with scores, speed and memory.
- [Hermes Agent v0.22.5 retest: how 12 local models changed after the upgrade](https://localagentboard.com/news/hermes-agent-v0-22-5-retest-local-models/): We reran our llama.cpp setups on Hermes Agent v0.22.5. Most scores moved by 2 points or less, within normal noise. Old vs new scores for every setup, and which ones are under review.
- [LocalAgentBoard launches: a tested leaderboard of local models for Hermes Agent](https://localagentboard.com/news/localagentboard-launch-local-hermes-agent-leaderboard/): LocalAgentBoard is an independent leaderboard that tests which local LLMs, quants and engines actually complete Hermes Agent tasks, measured on an NVIDIA DGX Spark. 16 setups tested so far.
- [Q4 vs Q8 for local agents: does quantization hurt Hermes Agent scores?](https://localagentboard.com/news/q4-vs-q8-quantization-local-agent-benchmark/): We tested Gemma 4 26B, Qwen3.6-35B-A3B and Nemotron-Cascade-2-30B at Q4, Q6 and Q8 on the same Hermes Agent tasks. Q4 scored within noise of Q8 while using much less memory and running faster.
- [Why ±2 points is a tie: reading local LLM agent benchmark scores](https://localagentboard.com/news/why-2-points-is-a-tie-local-llm-agent-benchmark/): The same local model can score 27, 30 and 28 out of 30 on identical agent tasks. Why LocalAgentBoard treats gaps of 2 points or less as a tie, and why Best overall is picked by speed within the top group.

## Model pages
- [Qwen3.6-27B Q6_K (GGUF) (llama.cpp)](https://localagentboard.com/models/qwen3-6-27b/): 30/30 tasks
- [Qwen3.6-35B-A3B UD-Q6_K (GGUF) (llama.cpp)](https://localagentboard.com/models/qwen3-6-35b-a3b/): 29/30 tasks
- [Qwen3.6-35B-A3B Q8_0 (GGUF) (llama.cpp)](https://localagentboard.com/models/qwen3-6-35b-a3b-q8/): 30/30 tasks
- [Qwen3.6-35B-A3B UD-Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/qwen3-6-35b-a3b-q4/): 29/30 tasks
- [Qwen3-Coder-Next FP8 (vLLM)](https://localagentboard.com/models/qwen3-coder-next/): not tested yet
- [Qwen3.5-122B-A10B NVFP4 (vLLM)](https://localagentboard.com/models/qwen3-5-122b-a10b/): not tested yet
- [Qwen3.6-35B-A3B NVFP4 (ModelOpt, safetensors) (vLLM)](https://localagentboard.com/models/qwen3-6-35b-a3b-nvfp4/): 30/30 tasks
- [gpt-oss-20b MXFP4 (official GGUF) (llama.cpp)](https://localagentboard.com/models/gpt-oss-20b/): 27/30 tasks
- [gpt-oss-120b MXFP4 (official GGUF) (llama.cpp)](https://localagentboard.com/models/gpt-oss-120b/): 30/30 tasks
- [Nemotron-3-Nano-30B-A3B Q8_0 (GGUF) (llama.cpp)](https://localagentboard.com/models/nemotron-3-nano-30b-a3b/): 25/30 tasks
- [Nemotron-Cascade-2-30B-A3B Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/nemotron-cascade-2-30b/): 29/30 tasks
- [Nemotron-Cascade-2-30B-A3B Q8_0 (GGUF) (llama.cpp)](https://localagentboard.com/models/nemotron-cascade-2-30b-q8/): 28/30 tasks
- [Nemotron-3-Super-120B-A12B NVFP4 (vLLM)](https://localagentboard.com/models/nemotron-3-super-120b-a12b/): not tested yet
- [Gemma 4 26B-A4B Q8_0 (GGUF) (llama.cpp)](https://localagentboard.com/models/gemma-4-26b-a4b/): 30/30 tasks
- [Gemma 4 26B-A4B UD-Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/gemma-4-26b-a4b-q4/): 30/30 tasks
- [Gemma 4 31B NVFP4 (vLLM)](https://localagentboard.com/models/gemma-4-31b/): not tested yet
- [Hermes-4.3-36B Q6_K (GGUF) (llama.cpp)](https://localagentboard.com/models/hermes-4-3-36b/): 22/30 tasks
- [Hermes-4-70B Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/hermes-4-70b/): 21/30 tasks
- [Hermes-4-14B Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/hermes-4-14b/): 13/30 tasks
- [Mistral Small 4 119B NVFP4 (vLLM)](https://localagentboard.com/models/mistral-small-4-119b/): not tested yet
- [GLM-4.7-Flash Q4_K_M (GGUF) (llama.cpp)](https://localagentboard.com/models/glm-4-7-flash/): 27/30 tasks
- [Qwopus-GLM-18B Q4_K_M (GGUF, healed) (llama.cpp)](https://localagentboard.com/models/qwopus-glm-18b/): 26/30 tasks