Open Models

Use auto, version-tracking aliases, fast tiers, or pinned Fireworks open models in coding workflows.

Use a Fireworks open model directly when you want Fireworks to serve every turn. FireConnect automatically registers the coding-ready catalog when you connect a harness.

Start with auto

auto chooses from frontier open models for each new user turn. It is the default on a standard Fireworks key.

Use auto-instant when you want a latency-first open-model mix. If fast models are disabled by model governance, use auto.

fireconnect claude

Restart Claude Code, open /model, and select Auto. No --model flag is required.

Choose a family, tier, or version

Use --model only when you want a specific family, tier, or version:

fireconnect claude --model glm-latest

Kind	Example IDs	Behavior
Latest	kimi-latest, glm-latest, deepseek-pro-latest	Tracks the current version in the model family
Fast	kimi-fast-latest, glm-fast-latest	Favors lower latency at a higher token price
Pinned	kimi-k3, glm-5p3, glm-5p3-flash	Stays on the named version

Prefer *-latest unless you need a pinned version.

Find a model ID

fireconnect model list
fireconnect model list --search glm
fireconnect model list --refresh

The catalog is cached for one hour. --refresh fetches the latest copy.

Check image support

glm-latest and glm-fast-latest are text-only. Image support for other aliases depends on the model currently behind each alias. Check fireconnect model list before sending images.

If an image reaches a text-only model in Claude Code, use /rewind or switch to a vision-capable model such as Kimi or glm-5p3-flash.

Use US-only models

Use the current model IDs, endpoint, and pricing guidance from US-only Serverless.

Pricing

Open models bill to your Fireworks account at Serverless Pricing. For per-session estimates and account usage, see Usage and Cost.