> ## Documentation Index > Fetch the complete documentation index at: https://docs.fireworks.ai/llms.txt > Use this file to discover all available pages before exploring further. # Models > Switch models in FireConnect harnesses: latest vs fast, when to pick each ID, and US-only routers ```bash theme={null} fireconnect model list --search kimi fireconnect on --model # Cursor / VS Code: quit the IDE before on, then reopen # Other harnesses: restart after on ``` `` is one of: `claude`, `opencode`, `codex`, `pi`, `cursor`, `vscode`, `deepagents`. There is no `model select` command. Always use `on --model`. ```bash theme={null} fireconnect opencode on --model kimi-fast-latest fireconnect claude on --model firerouter fireconnect claude on --interactive # Claude: map all six slots fireconnect claude on --opus glm-fast-latest --sonnet glm-fast-latest ``` Claude Code: `--model` sets **main** only; use `--opus`, `--sonnet`, `--haiku`, `--fable`, `--subagent`, or `--interactive` for the rest. Other harnesses take one `--model` for the default. Re-running `on` without `--model` keeps your current choice. | Harness | Apply the change | | -------------------------------- | --------------------------------------------------- | | Cursor, VS Code | **Quit the IDE**, run `on --model`, then reopen | | Claude Code | New session, or `/exit` then `claude --resume ` | | OpenCode, Codex, Pi, Deep Agents | Restart the CLI after `on` | ## Latest vs Fast FireConnect short IDs follow the same [Serverless serving paths](/serverless/serving-paths) as the API: | Kind | Example IDs | When to use | | -------------------------- | ------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | | **Latest** (standard path) | `kimi-latest`, `glm-latest` | Best price/quality. Version-tracking routers that move to the current model automatically. | | **Fast** | `kimi-fast-latest`, `glm-fast-latest` | Interactive coding where token speed matters. Same model quality as latest, higher \$/token, aims for **100+ tok/s**. | | **Pinned** | `kimi-k3`, `glm-5p2`, `kimi-k3-fast` | Stable ID that does not track new versions. | Prefer `*-latest` / `*-fast-latest` unless you need a pin. Browse live IDs and prices with `fireconnect model list`. ## Which model when | ID | Use when | Notes | | ------------------- | ----------------------------------------------------------------- | ------------------------------------------------------ | | `kimi-fast-latest` | Default interactive coding; screenshots / UI | Vision. Default for most harnesses and Claude `fable`. | | `kimi-latest` | Strong agentic coding with vision, lower \$/token than Fast | Vision. Standard path. | | `glm-fast-latest` | Fast text-only agent loops (Claude `opus` / `sonnet` defaults) | Text-only. 1M context. | | `glm-latest` | Cheaper text-only coding / long context | Text-only. Same family as Fast at standard price. | | `deepseek-v4-flash` | Haiku / subagent / high-volume background work | Text-only. Lowest cost in the coding catalog. | | `firerouter` | Auto-route easy work to open models, hard work to Claude Opus 4.8 | See [FireRouter](/ecosystem/firerouter/overview). | Also useful from `fireconnect model list`: `minimax-latest` / `qwen-plus-latest` (cheaper vision), `deepseek-v4-pro` (stronger text coding than Flash), `kimi-k2p7-code` / `kimi-k2p7-code-fast` (code-focused Kimi). GLM and DeepSeek Flash/Pro are **text-only**. Pasting images on those slots in Claude Code can break the session. Recover with `/rewind`, or use a Kimi ID. See [Claude Code troubleshooting](/ecosystem/fireconnect/claude-code#troubleshooting). ## US-only For US-only inference (compliance), pass the **full router path** with `on --model`. FireConnect does not yet expand the short US slugs to routers (unknown short IDs map to `accounts/fireworks/models/...`, which is wrong for these endpoints). | Model | Full ID | | ----------------- | -------------------------------------------- | | Kimi K3 (US) | `accounts/fireworks/routers/kimi-k3-us` | | GLM 5.2 Fast (US) | `accounts/fireworks/routers/glm-5p2-fast-us` | ```bash theme={null} fireconnect claude on --model accounts/fireworks/routers/kimi-k3-us fireconnect opencode on --model accounts/fireworks/routers/glm-5p2-fast-us ``` US-only endpoints are a **10% premium** except GLM 5.2 Fast US (same price as global Fast). Details: [US-only Serverless](/serverless/us-only-serverless). ## Limits * **Fire Pass** (`fpk_...`): Claude defaults to `kimi-fast-latest`. Catalog is limited. Not on Codex. FireConnect rejects `--model firerouter` with a Fire Pass key on **every** harness; use an `fw_...` account key for FireRouter. * **Foundry**: pass the Azure deployment name (`FW-GLM-5.2`), not a short ID. Claude Code does not support Foundry. * **FireRouter**: standard key only (`fw_...`), not Fire Pass. Cursor / Deep Agents need workspace BYOK. ## Troubleshooting Restart the harness. For Cursor and VS Code, fully quit the IDE **before** running `on`. FireConnect rejects `--model firerouter` when your stored key is Fire Pass (`fpk_...`), on every harness. Switch to an `fw_...` account key, then run `fireconnect on --model firerouter` and restart the harness. Pass the Azure **deployment name** (for example `FW-GLM-5.2`), not a serverless short ID. ## See also * [FireConnect overview](/ecosystem/fireconnect/overview) * [CLI reference](/ecosystem/fireconnect/cli-reference) * [Harness guides](/ecosystem/fireconnect/claude-code)