Models, free models and your own keys

Sign in to Vector to use free models, or connect a provider account or API key. The free models, their data policies and their limits are described below.

Bring your own key

Connect a provider in Settings, or run vector providers login in the terminal, then select a model in the session composer. Model availability, context limits, pricing, and rate limits come from the provider. Vector's usage view tracks the information returned by those sessions.

From 1.99.99, Vector no longer loads the former third-party hosted providers, even with an API key. Choose another provider before starting a task. Sign in with ChatGPT is off in Vector: connect OpenAI with an API key, or run your own installed Codex CLI from Host locally in the desktop app, where it signs in by itself. A ChatGPT sign-in saved earlier is no longer used. Native Copilot, xAI, GitLab OAuth, Poe and DigitalOcean sign-in flows are paused pending Vector's own registrations; API-key methods remain where supported. Claude Code, Codex, Cursor CLI and the other agents in Host locally still use your own installed commands and their own sign-in.

Free models inside of Vector

Sign in to Vector and the Free models inside of Vector are there at no cost to you: they are provided by Vector, on Vector's own OpenRouter account, through Vector's server, which checks your account and its limits. You don't need an OpenRouter account of your own. Every model is listed at a zero price with tool-capable endpoints, and a request never switches to a paid model or provider. The section appears while Vector's service has free models switched on and has models to serve.

From desktop 1.99.111 and the next CLI release, Vector offers every free model OpenRouter lists that can use tools, which Vector's agent works through. Google's Lyria previews and NVIDIA's content-safety classifier are left out because they list no tools, and a model whose free endpoints have been down all day returns once they answer again. A new free model appears on the next daily refresh. NVIDIA's free Nemotron models are included; NVIDIA's terms for them allow trial and evaluation use. Each model carries the data policy Vector recorded for the provider serving it, reviewed on 17 October 2026, and one Vector has not reviewed yet is treated as one that may train:

• No training: Cohere's North Mini Code (Cohere, kept up to 30 days), Dots3-Note Preview (AtlasCloud, which does not publish how long it keeps prompts) and XingChen's Xing4.0 (Phala). Each request asks OpenRouter to deny data collection. Google's Gemma 4 26B and 31B (Google AI Studio, kept up to 55 days) are also provided by Vector, with a small “US only” tag in model lists. This is a label only; Vector does not check your location.
• May train on your prompts: Poolside's Laguna S 2.1 and XS 2.1 (free until 31 October 2026), Liquid's LFM2.5-2.6B, Thinking Machines' Inkling and Inkling Small, NVIDIA's Nemotron models, OpenRouter's Free Models Router, and every free model Vector has not yet reviewed, such as Recursive's Flint. Their providers may log your prompts, code and outputs and use them to train their models. The Free Models Router sends each request to a free model OpenRouter picks at random.

The model picker, the terminal's model list and vector models show “May train on your prompts” on those models. Before your first prompt to one, Vector asks you once, and nothing is sent until you accept. That one acceptance covers every model with the label, including ones Vector has not reviewed yet, until you withdraw it. A prompt to one of them falls back only to another such model from the same provider, never to a different provider. An invited guest in a shared workspace is not offered these models, since only the workspace owner can accept the notice. Once a guest has worked in a session, nothing from that session goes to them, your own prompts in it included, because the guest's messages are part of what each request sends. In the terminal, vector models --accept-training accepts and vector models --decline-training withdraws; in the desktop app, Withdraw is on the Free models inside of Vector card in the Help panel. Vector itself never trains on your prompts. Vector never picks one of these models for you ahead of a no-training model. Desktop 1.99.108 and CLI 1.99.99 are offered no-training models only.

Free models provided by Vector have a per-minute and a daily limit for each account, and OpenRouter's own limits for Vector's account also apply. A coding task can use many requests, since each tool step is one. At a daily limit, Vector says when more free tokens are available, and at a per-minute limit when to try again. To keep working before then, choose a model from another provider you've connected, or add your own provider key in Settings → Providers.

Prompts are processed by OpenRouter and the selected model provider. Vector does not request zero data retention for these models: most of their providers keep prompts and outputs under their own policies, and OpenRouter records Phala, which serves Xing4.0, as keeping none. The policies rely on the providers' published terms and OpenRouter's records, not an independent audit. They do not cover Vector's local history, OpenRouter's account logging or other connected services. Review them before sending confidential code. Other provider connections have their own terms.

OpenRouter is also an ordinary provider in Settings → Providers. If you connect an OpenRouter account of your own there, its free models are listed under OpenRouter with that account's other models, each named with “(free)” after it, such as “Qwen3 Coder (free)”, and run on that account's own limits and privacy settings. The composer's model button shows the same name once one is chosen. Free models inside of Vector lists only the models provided by Vector, which keep running on Vector's account while yours is connected. Each row's tooltip says where it runs: Provided by Vector, or Your OpenRouter account. A key stored only in a repository's provider.openrouter.options.apiKey does not list them. From 1.99.99 to 1.99.108, connecting OpenRouter was the way to start free; from desktop 1.99.109 it is signing in to Vector. OpenRouter routing documentation describes the data-collection and provider controls.

Models in the terminal

vector models lists the supported model catalog. Connect a provider with vector providers login to use its models. Pass a provider id to filter and --verbose for metadata such as per-token cost. The catalog is bundled with the release, and from 1.99.99 Vector refreshes it from Vector's mirror, https://vectordev.ai/models, once it has loaded and then every hour, keeping the last catalog it had when the mirror cannot be reached. VECTOR_MODELS_URL points it at another mirror, VECTOR_MODELS_PATH selects a local catalog file and stops the refresh, and VECTOR_DISABLE_MODELS_FETCH=1 stops the refresh too. --refresh fetches at once, but only when VECTOR_MODELS_URL is set. With Vector Local on, the refresh is blocked unless you allow it or the mirror is on this computer.

vector models vector models anthropic --verboseCopy

Signed in with vector login, the free models provided by Vector are listed under Free models inside of Vector, marked Provided by Vector. Display names omit :free; selectors retain it. Human guidance, including the “May train on your prompts” label from the next CLI release, goes to stderr, while stdout keeps plain model selectors.

Models that expose reasoning effort take --variant. The level resolves from the flag, then your saved preference, then what the session last used, and your choice is kept per provider and model.