Ever had this happen? OpenRouter ships a new model with great price-to-performance and reviewers rave about its enormous context window — you open Hermes’ model picker, scroll through the list… nothing. You’re stuck waiting for the official catalog to catch up, or hand-rolling config to force it in. The PR merged on August 26 (#95621) adds another member to the picker: z-ai/glm-5.3-flash, live in both the OpenRouter and Nous Portal catalogs — and its headline feature is exactly what long-session users want: a 1,048,576-token (~1M) context window.

The New Model’s Profile

GLM-5.3-Flash is the “fast, big-window” member of Zhipu’s GLM-5.3 family. In the Hermes catalogs it slots below glm-5.3 and above glm-5.2 (newest-first ordering). Key specs (from OpenRouter live metadata):

Item	Value
Model ID	z-ai/glm-5.3-flash
Context window	1,048,576 tokens
Input price	$0.075 / M tokens
Output price	$0.25 / M tokens
Max output per call	131,072 tokens

Catalog totals after the update: 43 models on OpenRouter, 32 on Nous Portal.

How to Use It in Hermes

Three ways — interactive, CLI, or in-session:

Way 1: the interactive picker

hermes model

Follow the prompts to pick a provider (OpenRouter or Nous Portal) and the model. If you’ve opened the picker before and the cache predates the update, refresh first:

hermes model --refresh

--refresh wipes the picker’s disk cache and re-fetches each provider’s live /v1/models list — the way to see new models the moment they land.

Way 2: specify at startup

hermes -m z-ai/glm-5.3-flash

One-off sessions only; doesn’t change your default.

Way 3: switch in-session

Use the /model command to search and switch; typing glm-5.3 fuzzy-matches z-ai/glm-5.3-flash. Combined with the command palette’s /model filter, the swap takes seconds.

What 1M Context Actually Means

• Marathon sessions stop stressing: a 500K-token conversation uses half the window; combined with the new lean compaction default, compaction only kicks in near 900K — you barely think about it mid-run;
• Feed the whole repo to the agent: a large codebase fits as context instead of juggling files — a natural partner for unlimited max_turns;
• Do the math on cost: a big window is not free — tokens are metered, and a full 1M-token input costs ~$75 per shot. The rational pattern is to treat the window as a buffer, not a warehouse, and lean on compression to keep actual inputs in check.

A Note of Caution

In the catalog, glm-5.3-flash resolves its context metadata through a fuzzy key inherited from glm-5.3 (1,048,576), matching OpenRouter’s live data. If you’re unsure about the model’s real-world capabilities, run it on a small task before production. To fine-tune behavior (e.g. cap output length), the model_overrides guide explains per-model overrides for context_window, max_output_tokens, and friends.

Wrapping Up

GLM-5.3-Flash entering the catalogs is a genuinely useful option for long-session users: million-token context, sensible pricing, selectable right from the official picker. Model catalogs move fast — running hermes model --refresh every now and then to see what’s new is already part of the daily routine for Hermes power users.