Hermes Agent needs 64K context: what that means for qwen3 and Ollama

Hermes asks for 64,000 tokens; Ollama's default gives a small fraction of that

The rule comes straight from the Hermes docs. Checked 14 October 2026, the providers page states that agent use with tools needs at least 64,000 tokens of context, because the system prompt, tool schemas and working conversation state need room for reliable multi-step work. In practice Hermes Agent refuses a local window below 64K and will not run until it is raised.

The Hermes providers docs list the same default as Ollama's own documentation (docs.ollama.com/context-length, checked 13 October 2026): it depends on memory and is 4,096 tokens under 24 GB, the bracket a 16 GB Mac mini falls in. The 2,048 figure comes from a startup error message the docs list ("Context limit: 2048 tokens"); the older local-Ollama guide still repeats it. Either way the window is tiny next to the 64,000 minimum, which is what the chart shows.

Two settings clear the check: one on the Ollama server, one in the Hermes profile

These are the same two the Staffbox installer (scripts/install.sh) makes. Set the window where the model runs, then tell Hermes what it has.

1. Set the window on the Ollama server. The installer puts this line in the environment of the launch agent that runs ollama serve:

OLLAMA_CONTEXT_LENGTH=65536

2. Set the window in the Hermes profile. In the profile's config.yaml, the installer writes these keys in the model: block (the script's own comment says that without ollama_num_ctx Ollama falls back to a 2048-token window):

model:
  default: qwen3:8b
  provider: ollama
  base_url: http://localhost:11434/v1
  context_length: 65536
  ollama_num_ctx: 65536

Running Hermes for a business? Ask us

Alternative: a Modelfile variant

You can instead bake the window into a copy of the model. This is not what the installer does; we use it on internal test machines. Save as Modelfile:

FROM qwen3:8b
PARAMETER num_ctx 65536

Create the variant, then set default: qwen3-8b-64k in the Hermes profile instead of qwen3:8b:

ollama create qwen3-8b-64k -f Modelfile

Memory: Ollama's docs say a larger context window needs more memory. The memory cost at 64K has not been measured on our unit. Changing the window makes the model reload, which took about 100 seconds on our 16 GB Mac mini.

Every Mac mini score was measured with the window at 65,536

The installer settings above were in place for every Mac mini run we report, so no Mac mini score comes from a smaller window.

• Qwen3 8B with company notes and the quote action, 16 GB Mac mini M4: 15/15 on an unseen messy set, 40/40 (Fieldstone IT), 39/40 (Peachtree Cabinet Works), and 351/360 on a generated regression set (requests written by a model, answers checked by code), all on 6 October 2026. The companies are fictional. On a new trade it had never seen, an electrical distributor (14 Oct, Dell test box), the default Qwen3 8B setup scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55, because per-foot units, price levels and trade shorthand are not handled yet; that is why we score your own list first (the electrical result).
• The 27B model on the Dell RTX 3090 scored 40/40 and 39/40 in a run whose result file is named for the 64K setting (evals/results 2026-09-30, "27b-64k"). That is a different model on a different machine.
• Changing num_ctx costs about 100 seconds on the 16 GB Mac mini because the model reloads, so we keep one context size per run.

How we measured: fictional companies, answers graded by machine, one request at a time with the model warm, 16 GB Mac mini M4, 2 to 6 October 2026; results files on GitHub.

Prompts past qwen3's native 40,960 tokens are untested, so long-document jobs need a longer native window

The scores tell you how the unit behaves on the job it is sold for: short quote requests, far below 40,960 tokens. They say nothing about pasting a long contract or a whole mailbox into one prompt. That stretch, between the model's native window and the 65,536 we set, is the bracketed range on the chart, and we have no data on it.

If your workflow needs long prompts, say so before the test. The fix ladder is brain first, then a tool, then a bigger model, and a model with a longer native window is the last rung. Third-party sources report that gpt-oss (128K native) and newer Qwen releases with 256K clear the check (betterclaw.io, as reported, seen 13 October 2026). Our own results show gpt-oss-20b scoring 39/40, 40/40 and 15/15 on the Dell and needing 32 GB; its Apple speed is not measured. See the local models page and the non-Chinese model page.

If you are evaluating Staffbox, agree the pass bar before day 0 (see how to run an AI pilot), and see how the installer sets the window.

Sources

• Hermes docs: providers (64,000-token minimum and its reason; Ollama default 4,096 under 24 GB), checked 14 Oct 2026
• Hermes docs: local Ollama setup (64,000-token minimum), checked 14 Oct 2026
• Ollama docs: context length (default by memory: 4K below 24 GB; larger windows need more memory), checked 13 Oct 2026
• betterclaw.io: Hermes context length fix, as reported, seen 13 Oct 2026
• Hermes Agent issue 53347 (allow below 64K), as reported, seen 13 Oct 2026
• github.com/Staffbox-ai/staffbox (Staffbox code and evals/results, MIT), scripts/install.sh; evals/results 2026-10-02, 2026-10-06, 2026-09-30
• huggingface.co/spaces/staffbox/scorecard (every graded answer)