# Mistral Vibe + GLM-5.2 Setup How to run **Z.ai GLM-5.2** — the open-weights model Mistral has hosted on La Plateforme since 2026-08-11 — as the coding agent in **Mistral Vibe**, both in the CLI and the VS Code extension. Vibe is the agent formerly called Le Chat; its coding surface (`vibe for code`) ships as a CLI and a VS Code extension. GLM-5.2 is not a Mistral model — it is a third-party open model running on Mistral's infrastructure, under the same regional controls and service commitments as Mistral's own models. > **Why this page lives here.** The Kataglyphis repos are worked on by AI coding > agents. This is the setup guide for one of them, so any contributor (human or > agent) can reproduce the working config without reverse-engineering it from a > log file. The errors and fixes in the troubleshooting table were all found > live on 2026-08-28 — including two that produce no log entry, only garbled > TUI output. --- ## Prerequisites | Need | Where | |---|---| | Mistral Vibe CLI | `curl -LsSf https://mistral.ai/vibe/install.sh \| bash` or `uv tool install mistral-vibe`; on a native riscv64 host `bash linux/host-config/install-mistral-vibe-riscv64.sh` ([`linux-host-setup.md` § D4](linux-host-setup.md#d4-python-cli-tools-that-build-from-source-on-riscv64)) | | Mistral account + API key | [console.mistral.ai/api-keys](https://console.mistral.ai/api-keys) (La Plateforme / Mistral Studio) | | VS Code ≥ 1.94.0 | only if you want the IDE panel — the CLI works standalone | Verify the CLI is on PATH after install: ```bash vibe --version ``` ## 1. Get a Mistral API key 1. Sign in at [console.mistral.ai](https://console.mistral.ai). 2. Go to **API Keys** and create a new key. 3. Copy it immediately — it is shown once. A real Mistral key is roughly 32 alphanumeric characters. If what you have is shorter or reads like a placeholder (`DEIN…`, `your-key-here`), it is not a real key and every request will fail with `HTTP 401`. ## 2. Write `~/.vibe/config.toml` Vibe stores its config in `~/.vibe/config.toml` (Windows: `%USERPROFILE%\.vibe\config.toml`). The block below is the known-good config for GLM-5.2 on the EU endpoint: ```toml active_model = "zai-glm-5-2" [[providers]] name = "mistral-eu" api_base = "https://api.eu.mistral.ai/v1" api_key_env_var = "MISTRAL_API_KEY" api_style = "openai" backend = "mistral" [[models]] name = "zai-glm-5-2" provider = "mistral-eu" alias = "zai-glm-5-2" display_name = "Z.ai GLM 5.2" temperature = 1.0 thinking = "off" supports_images = false input_price = 1.4 output_price = 4.4 cached_input_price = 0.14 auto_compact_threshold = 800000 ``` ### Field reference | Field | Value | Why | |---|---|---| | `name` | `zai-glm-5-2` | The exact model ID the Mistral API expects. Not `glm-5.2`, not `glm5.2`, not `glm-5-2` — those all return `Invalid model`. | | `provider` | `mistral-eu` | Matches the `[[providers]]` block below. Use `mistral-eu` for the EU endpoint, or point `api_base` at `https://api.mistral.ai/v1` for the US endpoint. | | `backend` | `"mistral"` | Uses the official Mistral SDK backend, which maps Vibe thinking levels to the API's `reasoning_effort` parameter. The `"generic"` backend sends the raw level string (`"max"`) which the Mistral API rejects — it only accepts `"none"` or `"high"`. The default Mistral provider in Vibe ships with `backend = "mistral"`. | | `alias` | `zai-glm-5-2` | What shows in the model picker. Keep it identical to `name` to avoid `/model` confusion. | | `display_name` | `Z.ai GLM 5.2` | Human-readable label in the picker. Optional but useful when multiple models are configured. | | `thinking` | `"off"` | GLM-5.2's model card lists no reasoning mode. Setting any non-`"off"` value makes the backend send a `reasoning_effort` parameter the model does not support, producing garbled output in the TUI. | | `supports_images` | `false` | GLM-5.2 is text-only. | | `input_price` | `1.4` | Cost per million input tokens (USD). Lets Vibe track session cost. From the [model card](https://docs.mistral.ai/models/zai-glm-5-2). | | `output_price` | `4.4` | Cost per million output tokens (USD). | | `cached_input_price` | `0.14` | Cost per million cached input tokens (USD). | | `auto_compact_threshold` | `800000` | The model has a 1M-token context window. Compact at 800K to leave headroom. | | `active_model` | `zai-glm-5-2` | Must not be empty — an empty `active_model` means Vibe starts with no model selected. | ## 3. Set the API key The key is **not** stored in `config.toml`. Vibe reads it from the environment variable named by `api_key_env_var` (here `MISTRAL_API_KEY`). The easiest way to set it persistently is the built-in setup flow: ```bash vibe --setup ``` This opens a terminal prompt; paste the key from step 1. Vibe writes it to `~/.vibe/.env` and future sessions pick it up automatically. To verify the key is loaded without starting an interactive session: ```bash # quick smoke test — model replies and exits vibe -p "Reply with: OK" --max-turns 1 --auto-approve --output text ``` ## 4. VS Code extension (optional) The **Mistral Vibe VS Code** extension (`mistralai.mistral-vibe-code`, on the Marketplace) runs the same bundled agent as the CLI. It reads the **same** `~/.vibe/config.toml` and `~/.vibe/.env`, so once the CLI is configured the extension picks up GLM-5.2 with no extra setup. To install from inside VS Code: 1. `Ctrl+Shift+X` to open Extensions. 2. Search **Mistral Vibe**. 3. Install the one published by **Mistral AI** (`mistralai`). 4. Open the Vibe panel from the activity bar, or run **Mistral Vibe: Open Mistral Vibe** from the Command Palette. If the extension needs credentials and none are configured yet, it opens a terminal setup flow that runs `vibe --setup` — step 3 above. --- ## Troubleshooting Every row below was hit live on 2026-08-28. The log is at `~/.vibe/logs/vibe.log`. | Symptom (log message) | Cause | Fix | |---|---|---| | `Invalid model: glm-5.2` / `Model 'glm-5.2' is not available on mistral` | Wrong model ID in `config.toml` — the API rejects anything that is not exactly `zai-glm-5-2`. | Set `name = "zai-glm-5-2"` (and `alias` to match). | | `Invalid API key. Please check your API key and try again.` (on **every** model, including `mistral-medium-3.5`) | `~/.vibe/.env` contains a placeholder or stale key, or `MISTRAL_API_KEY` is unset. | Run `vibe --setup` and paste a real key from [console.mistral.ai/api-keys](https://console.mistral.ai/api-keys). | | `HTTP 401` at startup (`Admin-managed config not applied` / `Connector catalog is unavailable`) | Same root cause — the bad key fails before any model call. | Same fix — set the key. These warnings clear once auth works. | | **Garbled / strange output in the TUI** (no log entry; the model answers but the display is wrong) | `backend = "generic"` sends `reasoning_effort: "max"` raw — the Mistral API only accepts `"none"` or `"high"`. The `"mistral"` backend maps levels correctly (`"max"` → `"high"`), but GLM-5.2 has no reasoning support so the parameter should not be sent at all. | Set `backend = "mistral"` **and** `thinking = "off"`. With thinking off, no `reasoning_effort` is sent regardless of backend. | | Startup warning about `thinking` / reasoning mode | `thinking = "max"` set for a model whose card lists no reasoning mode. | Set `thinking = "off"`. Vibe's own migration code does the same for `devstral-small-latest` (`"Force thinking='off'; it has no reasoning support"`). | | No model selected on startup; `/model` shows nothing active | `active_model = ""` in `config.toml`. | Set `active_model = "zai-glm-5-2"`. | | Model not in the picker at all | `[[models]]` block missing or `provider` does not match any `[[providers]]` `name`. | Ensure `provider = "mistral-eu"` matches the provider block, and the model block is a top-level `[[models]]` table. | ### Quick health check ```bash # 1. is the CLI installed? vibe --version # 2. is the key set in this shell? (prints its length only, never the key; # 0 is fine when vibe --setup stored it in ~/.vibe/.env instead) echo "${#MISTRAL_API_KEY}" # 3. does the model answer? vibe -p "Reply with: OK" --max-turns 1 --auto-approve --output text ``` If step 3 prints `OK`, the full chain — key, endpoint, model ID — is correct. --- ## References - **GLM-5.2 model card** (features, context, pricing): - **Mistral Vibe product page**: - **VS Code extension — install & auth**: - **Vibe CLI configuration**: - **Reasoning on Mistral** (`reasoning_effort` parameter, supported models): - **GLM-5.2 on Mistral announcement** (2026-08-11): - **API keys**: