Hermes Agent Cursor + Which AI Models To Run After You Import: Real GoldieBench Scores For Cursor Users

Hermes Agent Cursor import moves your rules, MCP servers and skills into Hermes, but it never brings a model with it, so the first choice after the import is which model to run.

The importer, hermes import-agent cursor, comes from Nous Research's agent-importers plugin and never reads ~/.cursor/cli-config.json.

Below I match models to the jobs Cursor users usually bring across, using only scores I read on live GoldieBench pages on 1 November 2026.

If you want my team to help you plan the agent side of your marketing, book a free SEO strategy session.

Why the model question comes straight after a Cursor import

The Hermes import guide is clear that API keys and credentials are never imported.

Your memory, MCP servers and skills arrive; the model you used inside Cursor does not.

You pick a provider again with hermes setup.

That makes the import a clean moment to choose on scores rather than habit.

Your imported rules sit in MEMORY.md, your MCP servers in config.yaml, and your skills in a cursor-imports folder.

Whichever model you choose reads all three on every session.

What GoldieBench actually measures

Per the methodology page, every model gets the same fixed prompt, one shot, and returns a single HTML file.

Julian scores each result from 0 to 10, and the leaderboard averages those scores across the tasks each model ran.

Tasks fall into Games, Visuals, Sims and Pages, and each model row shows what it is best at.

That is building, which is close to what most Cursor users did all day.

The scores, as read on goldiebench.com

The local score comes from the separate local leaderboard, where Qwable 5 27B Coder ranks first.

Context figures come from each model page; "n/a here" means I did not read one for that row.

A main model for day-to-day Hermes work

For the coding and page-building work most Cursor setups were made for, start with Claude Opus 5 or GPT-5.6 Sol.

Opus 5 sits at 8.27 across 50 tasks with 13 golds, the highest average in the table above.

GPT-5.6 Sol is 8.16 across 50 tasks and is the board's "best at Pages" pick among the top four.

If your Cursor rules are mostly about web pages and copy, Sol lines up closest with that work.

MoA costs more calls per answer than one model, because every panel member answers before the aggregator writes.

Keep it for the jobs where a better first draft saves you a rework.

Test MoA on one build where Opus 5 alone left you editing, and keep it only if the first draft needs less work.

A long-context model for the repos you brought from Cursor

Imported memory, MCP tools and skills all take up context, and a big codebase on top fills a window fast.

Kimi K3's page lists a 1,048,576-token context, GPT-5.6 Sol's lists 1,050,000, and Opus 5's lists 1,000,000.

Kimi K3 scores 7.89 over 50 tasks, and its page describes it as tuned for long-horizon agent work.

For long sessions on one repo, I would put Kimi K3 or Sol in the main slot.

GLM-5.2 is the other "best at Pages" model in the top half of the board, at 7.77 over 47 tasks.

Grok sits at 8.09, with 43 of its 47 tasks scored and "best at Games" on its row.

A model just for the sync job and other cron runs

The import guide suggests running hermes import-agent --sync on a daily cron job.

Hermes lets every unpinned cron job follow cron.model, separate from your chat model.

A sync report does not need a frontier model, so a lighter pick fits.

Gemini 3.6 Flash scores 7.08 over 50 tasks with a 1,000,000-token context, which is plenty for summarising what changed.

Set it once with hermes config set cron.model <name>.

Pinning matters here: a pinned job keeps its own model even when you change the main one, so pin the sync job if you ever move cron to something heavier.

A local model if your Cursor work must stay on your machine

The local leaderboard ranks Qwable 5 27B Coder first at 7.14 over 41 tasks.

It runs on Apple MLX and is listed at 15.0GB.

That is the strongest on-device score on the bench, and it sits above the 7.08 Gemini 3.6 Flash scores on the main board.

Pick it when client code cannot leave the machine.

Check the local board again before you commit, because models still benching show as pending there.

Hermes MoA: an option that only exists inside Hermes

Hermes MoA sends one prompt to a panel of models in parallel and has an aggregator write one final answer.

Its page lists 8.17 over 47 tasks with "best at Visuals", and a default panel of Claude Opus 4.8 and GPT-5.5.

It is a system, not a single model, and it only runs inside Hermes.

For high-stakes builds after the import, it is the setup worth testing against Opus 5.

How to switch models after the import

1. Run hermes setup to connect a provider, since no keys came over from Cursor.
2. Use hermes model to set the main chat model.
3. Set cron.model for the sync job and other scheduled runs.
4. Pin one job to its own model with hermes cron edit <job_id> --pin when it needs something heavier.

A simple pairing for most Cursor users

Start with that table, then swap one slot at a time and compare the output on your own jobs.

Changing the main model never touches the cron fleet once cron.model is set, so you can test freely.

What the scores do not tell you

GoldieBench grades one-shot builds, not how a model follows imported memory over a long session.

No task tests MCP tool calls or rule-following in Hermes, so treat these as a strength signal, not a guarantee.

Run one real Cursor job through your top two picks before you commit, and see my Hermes Import Agent explainer for the Claude Code and Codex side.

For the full Agent OS with models already wired, join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step run order for importing a Cursor setup into Hermes
• AI Income Deskwhether a small business should move its Cursor setup into Hermes
• AI Tool Verdicta scorecard verdict on the Cursor import, rated 8.3 out of 10
• agentos.guidefive copy-paste prompts that get your agent importing Cursor

Put Opus 5 or Sol in the main slot, give cron a lighter model, and if you want help, book a free SEO strategy session or join the AI Profit Boardroom to run Hermes Agent Cursor.