Jan review

Jan is a desktop application for chatting with local models or connecting to online model providers. The homepage frames it as a way to choose models and use them through Jan Desktop, rather than as a hosted API, library, CLI or self-hosted server. Its free plan includes unlimited local inference and model downloads, so the local model runs on the user’s machine; online models use the user’s own provider API keys.

In a stack, Jan sits at the desktop-client layer: it gives an individual user a place to select a model and chat, while local compute or a third-party provider handles inference. The homepage lists OpenAI, Anthropic, Google, Meta, Mistral, Alibaba, DeepSeek and Moonshot AI. Jan advertises a Mac download. That makes it a fit for evaluating models or personal workflows, not a drop-in backend for an application. Memory features are planned, but the homepage labels them “Coming Soon.”

The homepage links to Jan’s docs and presents the product as a desktop app for choosing local or online models. It names no Jan API style, endpoint, SDK, programming language, CLI, webhook or plugin. The integration surface it does name is model selection in Jan Desktop and online-provider connections using your own API keys.

That makes Jan straightforward to assess as a user-facing client, but leaves code-level integration an open question. If you want to call Jan from an application, automate it in a deployment pipeline or manage it across a team, confirm SDK and API support in the vendor’s docs before committing to that architecture. Do not assume that the providers Jan lists expose their own APIs through Jan in a form your code can call. The homepage also links to research and company information, and says Jan is built in public, with v0.8.8 announced as live on GitHub.

The pricing information lists one plan: Free / Open Source at $0 per month. It includes unlimited local inference, unlimited model downloads and an Apache-2.0 license. Jan has no paid tiers, and no annual billing price is listed. For local use, Jan does not charge by request, seat or inference volume under this plan.

Online model connections are a separate cost. Jan uses your own API keys, and third-party providers bill you directly. The total therefore depends on the provider and your usage; the pricing information gives no per-token or request rates for those services. There is no Jan overage charge listed, but provider charges can still grow as usage increases. Local inference also uses your computer’s resources, rather than a Jan-hosted compute plan. Teams should budget the provider’s rates separately and check that provider’s limits and billing terms. The free plan is the only Jan plan listed.

Jan describes itself as free and open source, and the free plan lists the Apache-2.0 license. The homepage says development is public and announces a release on GitHub. Its local inference option lets users run models on their own machines; this is desktop use, not a documented Jan server deployment.

That local option reduces dependence on Jan-hosted inference. For online models, charges and API keys belong to the third-party provider, so those connections are not a Jan-managed cloud account. Jan supports choosing among open and online models, which offers some flexibility at the client layer. The vendor does not publish a data-export workflow on the homepage; check the docs before relying on Jan to store conversations or preferences you need to move. Memory is still planned, so avoid making it a required part of a current workflow.

Jan’s free plan lists unlimited local inference and model downloads, but the vendor does not publish rate limits, service-level commitments or scaling guidance on the homepage. Local performance depends on the machine running the model; online use also depends on the selected provider and its limits. Before standardizing on Jan, check the vendor’s docs for security, compliance and operational details that matter to your deployment.

Reddit reports are mixed and need context. In one thread, Jan’s team said a sponsored download promising image generation was unaffiliated; commenters reported Trojan warnings and a broken GitHub link. That is a warning about an impersonating download, not evidence about the official app. In a separate LocalLLM thread, a user described running Jan from external storage, then later said the setup broke. Other commenters questioned the setup and noted that model loading and hardware affect the experience. Treat those anecdotes as caution, not a reliability measure.