Top AI Models for Hermes Agent (Tier List)

After a month of running Hermes Agent across nearly every major model, the Boxmining team ranks GPT 5.4 and Gemini 3.1 Pro as the top orchestrators, with GLM 5.1 and Kimi 2.5 owning the executor slot.

Orchestrators (the brain)

GPT 5.4 is the new king, replacing nerfed Claude Opus 4.6. It was "originally designed with native agentic workflows in mind," and Ron says Michael has been "pretty happy with the output" from his Hermes agent on it. Gemini 3.1 Pro ties as top orchestrator but adds native video and audio input, making it the go-to for screen recordings and structured dashboard extraction. Qwen 3.6 Plus earns the third orchestrator slot because its chain-of-thought stays active on every response — there's no toggle, and reasoning persists across prior turns, which cuts contradictions on long-horizon tasks. Kimi 2.5 rounds it out as an orchestrator-capable model thanks to its swarm agents feature, which can "self-direct a swarm of like about 100 sub agents" and coordinate up to 1,500 tool calls without a predefined workflow.

Executors (the hands)

GLM 5.1 is the standout — it one-shotted a space shooter benchmark that Opus 4.7 choked on, and survives Hermes' 85% auto-compaction events with strong context recovery, something Claude models failed at. Switching the entire agent fleet is a one- to two-line edit in config.yml. Minimax M2.7 is a strong executor (not orchestrator) because it was "trained on the Open Claw Agent Harness framework," the same lineage as Hermes, and Xiaomi is an official News Research Team partner. DeepSeek 3.2 earns executor for its ability to "think inside tool calls," eliminating redundant reasoning passes in Hermes' 40+ tool environments and slashing cron job errors. Nemotron 3 Super and Step 3.5 Flash fill remaining executor slots — both open-weight, with Nemotron self-hostable behind a Nemo Claw privacy wrapper.

Question mark: Claude

Opus 4.6, Opus 4.7, and Sonnet 4.6 all sit in question mark. On April 17, the same day Opus 4.7 dropped, the team ran it against GLM 5.1 on a simple one-shot space shooter build — GLM won, Opus "couldn't make like it broke." Sonnet 4.6 stopped following explicit HTML formatting instructions. Ron says compute is being shifted to Mythos, the next super model, but for now the recommendation is to hot-swap Claude out for the orchestrator and let Mimo V2 Pro handle execution.

Auxiliary (support only)

Gemini 2.5 Flash is the default baked into Hermes — check nano config.yaml. Gemini 3 Flash adds free Google Search grounding and URL context reading, a replacement for the $10/month News Research built-in tools. Mimo V2 Pro is the high-volume king and the most-used model on Open Router this month; free via the News API on the news portal website. Elephant Alpha (100B params, 256K context) and Trinity Large Preview are open-weight niche picks.