​
Model inclusion and lifecycle policy for the Venice API

The Venice API exists to give developers unrestricted private access to production-grade models free from hidden filters or black-box decisions. As models improve, we occasionally retire older ones in favor of smarter, faster, or more capable alternatives. We design these transitions to be predictable and low‑friction.

​
Model Deprecations

We know deprecations can be disruptive. That’s why we aim to deprecate only when necessary, and we design features like traits and Venice-branded models to minimize disruption. We may deprecate a model when:
  • A newer model offers a clear improvement for the same use case
  • The model no longer meets our standards for performance or reliability
  • It sees consistently low usage, and continuing to support it would fragment the experience for everyone else

​
Deprecation Process

When a model meets deprecation criteria, we announce the change with 30–60 days’ notice. Deprecation notices are published via the changelog and our Discord server. When you call a deprecated model during the notice period, the API response will include a deprecation warning. During the notice period, the model remains available, though in some cases we may reduce infrastructure capacity. We always provide a recommended replacement, and when needed, offer migration guidance to help the transition. After the sunset date, requests to the model will automatically route to a model of similar processing power at the same or lower price. If routing is not possible for technical or safety reasons, the API will return a 410 Gone response. If a deprecated model was selected via a trait (such as default_code, default_vision, or fastest) that trait will be reassigned to a compatible replacement. We never remove models silently or alter behavior without versioning. You’ll always know what’s running and how to prepare for what’s next.
Performance-only upgrades: We may roll out improvements that preserve model behavior while improving performance, latency, or cost efficiency. These updates are backward-compatible and require no customer action.
See the Model Deprecation Tracker below. For earlier announcements, consult the changelog and our Discord server.

​
How models are selected for the Venice API

We carefully select which models to make available based on performance, reliability, and real-world developer needs. To be included, a model must demonstrate strong performance, behave consistently under OpenAI-compatible endpoints, and offer a clear improvement over at least one of the models we already support. Models we’re evaluating may first be released in beta to gather feedback and validate performance at scale. We don’t expose models that are redundant, unproven, or not ready for consistent production use. Our goal is to keep the Venice API clean, capable, and optimized for what developers actually build. Learn more in Model Deprecations and Current Model List.

​
Versioning and Aliases

All Venice models are identified by a unique, permanent ID. For example: venice-uncensored qwen3-235b llama-3.3-70b deepseek-r1-671b Model IDs are stable. If there’s a breaking change, we will release a new model ID (for example, add a version like v2). If there are no breaking changes, we may update the existing model and will communicate significant changes. To provide flexibility, Venice also maintains symbolic aliases — implemented through traits — that point to the recommended default model for a given task. Examples include:
  • default → currently routes to llama-3.3-70b
  • default_code → currently routes to qwen-2.5-coder-32b
  • default_vision → currently routes to mistral-31-24b
  • default_reasoning → currently routes to deepseek-r1-671b
Traits offer a stable abstraction for selecting models while giving Venice the flexibility to improve the underlying implementation. Developers who prefer automatic access to the latest recommended models can rely on trait-based aliases. For applications that require strict consistency and predictable behavior, we recommend referencing fixed model IDs.

​
Beta Models

We sometimes release models in beta to gather feedback and confirm their performance before a full production rollout. Beta status does not guarantee promotion to production. A beta model may be removed if it is too costly to run, performs poorly at scale, or raises safety concerns. Beta models can change without notice and may have limited documentation or support. Models that prove stable, broadly useful, and aligned with our standards are promoted to general availability. To request early access, join us on Discord and let us know why you’d like to join the beta tester group.

​
Feedback

You can submit your feedback or request through our Featurebase portal. We maintain a public changelog, roadmap tracker, and transparent rationale for adding, upgrading, or removing models, and we encourage continuous community participation.

​
Model Deprecation Tracker

The following models are scheduled for deprecation. We recommend migrating to the suggested replacements before the removal date.
Deprecated ModelReplacementRemoval byStatusReason
deepseek-r1-671bqwen3-235bSep 22, 2025AvailableBetter model available, low usage
llama-3.1-405bqwen3-235bSep 22, 2025AvailableBetter model available, low usage
dolphin-2.9.2-qwen2-72bvenice-uncensoredSep 22, 2025AvailableBetter model available, low usage
qwen-2.5-vlmistral-31-24bSep 22, 2025AvailableLow usage
qwen-2.5-qwq-32bqwen3-235b (disable thinking)Sep 22, 2025AvailableLow usage
qwen-2.5-coder-32bqwen3-235bSep 22, 2025AvailableLow usage
deepseek-coder-v2-liteqwen3-235bSep 22, 2025AvailableLow usage
pony-realismlustify-sdxlSep 22, 2025AvailableBetter model available
stable-diffusion-3.5qwen-imageSep 22, 2025AvailableLow usage
flux-devqwen-imageOct 22, 2025AvailableBetter model available
flux-dev-uncensoredlustify-sdxlOct 22, 2025AvailableBetter model available