Model Guide
Hermes Agent + GLM 5.3 FlashX: The Same Flash, Up to 5x Faster
On 18 September 2026 Zhipu released GLM 5.3 FlashX — not a new model, but GLM 5.3 Flash served on infrastructure tuned for throughput, at up to 200 tokens/s. This page covers the Hermes Agent side: which provider slot it uses, the exact model code, and the price decision the extra speed asks you to make.
What you are actually selecting
FlashX is the same mixture-of-experts model as Flash: roughly 320 billion total parameters and about 18 billion active per token, natively multimodal — text, images and video in, text out — with a 1M-token context window and a 128K maximum output. Zhipu published no new benchmark numbers for FlashX, because the weights did not change.
What changed is the serving layer. Zhipu says FlashX reaches up to 200 tokens/s, about five times Flash's output speed, running on the same Chinese-chip inference cluster it has been scaling. So when you pick FlashX in Hermes you are buying time-to-first-finished-answer, not a smarter model.
Trap 1: the provider is called zai
Hermes names providers after the vendor that serves the model, not after the model family. Z.ai is the vendor; GLM is the family. So the provider id is zai while the environment variable is GLM_API_KEY. People grep their config for “glm”, find nothing, and conclude the provider is missing.
# ~/.hermes/.env
GLM_API_KEY=your-key-herehermes model
# choose z.ai / GLM, then GLM 5.3 FlashXIf Hermes is not installed yet, start with the installation guide.
Regional endpoints
To move to a different host, override the base URL instead of reaching for a custom provider:
# Standard API key
GLM_BASE_URL=https://api.z.ai/api/paas/v4
# GLM Coding Plan key — a different, dedicated endpoint
GLM_BASE_URL=https://api.z.ai/api/coding/paas/v4Set the override only if you actually need it. Leaving GLM_BASE_URL unset lets the provider resolve its own default, and hardcoding the general endpoint is precisely what breaks a Coding Plan key — the symptom is an authentication or model-not-found error against a key that works elsewhere.
Trap 2: the model code, and the route that does not exist yet
On Z.ai's own API the model code is glm-5.3-flashx, alongside glm-5.3-flash on the same documentation page. Aggregators lag a launch, and this one is a case in point: at publication time OpenRouter listed z-ai/glm-5.3-flash but not a FlashX slug. If you route through an aggregator, copy the identifier from that catalog rather than guessing the spelling, and do not assume the Flash slug will silently resolve FlashX.
| Route | Credential | Model ID |
|---|---|---|
| Built-in Z.ai provider | GLM_API_KEY | glm-5.3-flashx (selected through hermes model) |
| OpenRouter | OPENROUTER_API_KEY | Not listed at publication — check the live catalog |
The price decision
FlashX is priced at 2.5x Flash, which Zhipu presents as the cost of the extra speed. Published Z.ai rates per million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GLM 5.3 FlashX | $0.37 | $0.075 | $1.25 |
| GLM 5.3 Flash | $0.15 | $0.03 | $0.50 |
Same answers, 2.5x the token bill. That is a good trade when a human or another agent is waiting on each turn — a support bot, an interactive coding session, a tool loop where every second of latency is multiplied by ten steps. It is a bad trade for overnight batch work where throughput, not latency, decides the outcome.
Verify it is actually being used
After selecting the model, send one message and confirm the reply is attributed to FlashX rather than a fallback. Hermes will keep answering on a previously configured provider if the new key fails to load, and a working agent is not proof that your change took effect.
On OpenClaw Launch you can skip the provider plumbing: deploy a managed Hermes Agent, save your Z.ai key, and pick the model from the dashboard. One caveat: the picker lists what the provider catalog exposes, so GLM 5.3 FlashX appears there once Z.ai (via models.dev) or OpenRouter lists it. Until then the same base model is selectable as GLM 5.3 Flash, at 2.5x less. If you need FlashX today, the self-hosted config above is the reliable route.
Hermes Agent and GLM 5.3 FlashX FAQ
How do I set GLM 5.3 FlashX as the Hermes Agent model?
Put GLM_API_KEY in ~/.hermes/.env, then run hermes model and pick the Z.ai provider followed by GLM 5.3 FlashX. The provider id is zai — searching your config for “glm” will not find it.
Is GLM 5.3 FlashX a different model from GLM 5.3 Flash?
No. FlashX is the same 320B / 18B model as GLM 5.3 Flash, served on infrastructure tuned for throughput. Zhipu announced no new benchmark or architecture change — the difference is speed (up to 200 tokens/s, about 5x Flash) and price (2.5x Flash).
What is the GLM 5.3 FlashX model code?
The Z.ai API model code is glm-5.3-flashx. OpenRouter had not listed FlashX at publication time — its catalog carried z-ai/glm-5.3-flash but not the FlashX variant — so the direct Z.ai route is the one to use today. Check the live catalog before relying on an aggregator ID.
Does GLM 5.3 FlashX work with the GLM Coding Plan?
Not yet. Z.ai's documentation states the plan carries GLM 5.3 Flash (at three times the GLM 5.3 quota) but that GLM 5.3 FlashX is not yet available on the plan. On a Coding Plan, use Flash; reach for FlashX on pay-as-you-go until the plan adds it.
Is 2.5x the price worth it for the speed?
Only if latency is what is slowing you down. At $0.37 / $1.25 per million input / output tokens against Flash's $0.15 / $0.50, FlashX costs 2.5x for the same answers. It pays off in high-frequency agent loops and interactive sessions where a few seconds per turn compound — and not in overnight batch jobs where nobody is waiting. If a cheaper model is looping on a task it cannot finish, FlashX just makes the loop faster; fix the loop first.
Can I turn off thinking on GLM 5.3 FlashX?
No. Reasoning is mandatory on the Flash family. The Z.ai API accepts only thinking.type: “enabled”, so budget reasoning tokens as output when you compare FlashX against a model where thinking can be disabled.