Running Vibe with a Local Model (vLLM)

Vibe Mistral CLI, released this week, is Mistral’s take on the agentic code CLI.
However, one small detail is not documented anywhere: how to make Vibe work with local models instead of Mistral official APIs. While there is an option local, it will only work for localhost. I however have a dedicated GPU server that I want to integrate.

In this blog, I’ll walk you through how I successfully connected Vibe CLI to a locally hosted model running on vLLM, using Devstral-2-123B-Instruct-2512 as the example.

This makes it possible to use Vibe fully offline.

🌐 Background

Mistral released Devstral-2 and created the Vibe CLI tooling.
Their announcement mentions support for custom providers but does not include instructions for configuring local inference endpoints.

Reference release post:
https://mistral.ai/news/devstral-2-vibe-cli

Fortunately, Vibe internally speaks the OpenAI API protocol, which means it can connect to any server providing that interface — including vllm.

🛠️ Step 1 — Run your model with vLLM

On my setup, I’m running Devstral-2 on a GPU server that has less VRAM than the “official” 4×H100 recommendation.
I was forced to lower the context size to make it possible to fit my GPUs. vLLM makes this possible with memory-tuning flags.

Here’s the exact command I use:

vllm serve Devstral-2-123B-Instruct-2512   --tool-call-parser mistral   --enable-auto-tool-choice   --tensor-parallel-size 4   --max-model-len 128000   --gpu-memory-utilization 0.9

A few notes on these flags:

--tool-call-parser mistral
Keeps tool-calling compatible with how Devstral/Vibe expect function calls to be formatted.
--enable-auto-tool-choice
Lets the model automatically decide when to call tools instead of hard-coding them.
--tensor-parallel-size 4
Splits the model across 4 GPUs so you don’t need a single massive card.
--max-model-len 128000
Bumps the context window while still staying within the practical limits of your VRAM.
--gpu-memory-utilization 0.9
Pushes each GPU close to its memory limits so the model fits on smaller hardware.

Once running, vLLM exposes an OpenAI-compatible endpoint at:

http://YOUR-IP:8000/v1

This is exactly what Vibe expects.

📝 Step 2 — Configure Vibe to use your local model

Install and run Vibe CLI, this will create a configuration file:

~/.vibe/config.toml

And add:

[[providers]]
name = "vllm"
api_base = "http://some-ip:8000/v1"
api_key_env_var = ""
api_style = "openai"
backend = "generic"

[[models]]
name = "Devstral-2-123B-Instruct-2512"
provider = "vllm"
alias = "vllm"
temperature = 0.2
input_price = 0.0
output_price = 0.0

Explanation

api_base — Points Vibe to your local vLLM server
api_key_env_var = "" — vLLM does not require authentication
api_style = “openai” — tells Vibe to speak OpenAI-formatted requests
backend = “generic” — required for non-Mistral providers
input/output pricing — set to zero for local usage
alias — lets you call your model simply as vllm inside Vibe

▶️ Step 3 — Run Vibe with your local model

Now you can select the model in Vibe:

/model

And you now should be able to see vllm as a provider.

Vibe model select screenshot

🎉 Result

With this small configuration file, Vibe works perfectly with vLLM or could work with any other OpenAI compatible endpoint.

Running Mistral Vibe with a Local Model