Changelog

Securely Backup Chat History

Pro users can how securely backup chat history and migrate to other devices, or recover from a loss of data in their local browser. Here's how they work:

1. On your local device, when you create a new backup, Venice encrypts your data with a password that only you control.

2. That backup is then broken into chunks and uploaded to Venice's infrastructure.

3. You can then download and restore that backup, either overwriting your existing history, or merging it, on any logged in device.

A few important notes:

• Venice does not have any record of the password you create, so if you loose it, your backup is unrecoverable. We suggest you use a password manager to store them.

• You are limited to a max of 5 concurrent backups.

• Backups expire after 90 days.

Backups can be accessed via the menu in the left hand side menu:

From there, you can create a new backup, or restore an existing one:

App

• Maintain EXIF data on upscaled images.

• EXIF data on image generation now includes the model name.

• Updated default values for Dolphin 72B and adjusted additional inference settings based on recommendations from Cognitive Computations.

• Changing the conversation type selector will now change the image settings, but changing the image settings won’t automatically change the conversation type selector.

Mobile Wallets

• Mobile wallets (Coinbase Wallet, Metamask, etc...) will be redirected to the sign-in when visiting the Venice home page. This should reduce friction of logging in from those devices.

API

• Added model specific defaults for temperature and top_p - Updated the /models endpoint to list those defaults in the constraints field.

• Add support for the following parameters in the chat/completions endpoint:

  • repetition_penalty

  • max_temp

  • min_temp

  • top_k

  • min_p

  • stop_token_ids

Bug Fixes

• Fixed a bug where you could not click the scroll bar next to the chat input.

Venice is Burning

Venice is burning. 1/3 of our VVV token supply was burned, and we’ve rebranded.

Characters

• Fix an issue where character images crated with the character generator that were blurred were not showing the Safe Venice description on top.

• When editing a character, if you change the model and the context exceeds the model's new context, you'll now see an error presented on the context screen. Additionally, if you send a message that exceeds the context of the model server side, you'll get a character specific error directing you to change the character context.

• Improved context processing when conversations are nearing the maximum context of the mode.

API

• Support null value in Chat Completions stop parameter. Docs are updated.

• Overhauled significant portion of swagger documentation for Chat Completions API to make parameters and responses more clear in the docs.

App

• Updated the Image Settings “steps” tooltip to be more clear.

• Updated the UI so Reasoning models that never close the </thinking> tag will open the thinking content when the rendering is complete.

• Adjusted rate limits for users to debit only for successful requests.

• Venice Sticker factory prices have been reduced to $9.99.

• Implemented streaming of Venice Voice responses from our Venice Voice servers back to the client to reduce time to first speech.

• Add “Reasoning” feature to model dropdown for Reasoning models.

• Rewrote the app loading screen to remove flashes and other glitches during initial load. Introduce a smooth fade during this transition.

• Update the context length descriptions on our models to be more clear on available context within app.

• Added a warning when clicking links generated from LLMs.Added a warning when clicking links generated from LLMs.

Bug Fixes

• Fixed an issue with the sign out function occasionally requiring multiple calls.

• Fixed a bug where copying a WebP image using the contextual menu wouldn’t put the image on the clipboard.

• Fixed a bug with Safe Venice overlays appearing on non-adult images in some circumstances.

• Fixed a bug where under certain circumstances the user session token would not be refreshed before it expired. This would result in a screen suggesting the user’s clock was out of date.

• Fixed a UI quirk with API Key expiration dates where Safari would show a default date on the expiration selector despite the field being empty.

Characters with Pro Models accessible to all users

Characters that use Venice Pro models are now accessible for non-pro members to interact with. Anonymous and Free users will get a limited number of chats with these characters before being prompted to upgrade to Venice Pro. We look forward to your feedback on Venice Characters.

Venice Voice Downloads

Venice users can now download audio generate from Venice Voice. Once the audio has completely generated, a download icon will appear to the right of the speaker. Clicking this will allow you to save the audio recording as a .wav file.

App

• Refactored the Venice Voice UI so that the button in the message rows only controls reading for that particular message. For users who wish to have the whole conversation read, that can one enabled in the Text settings.

• Venice Voice pronunciation and processing was improved to provide better pronunciation and to strip out characters that are not processable.

• Fixed a bug where a user who was speaking to a character, then went to an image conversation and returned to the character would get an image generation in the first message with that character.

API

• Vision models will now support the submission of multiple image_url parts. For compatibility purposes, the schema supports submitting multiple image_url messages, however, only the last image_url message will be passed to and processed by the model. Docs have been updated.

• The model list endpoint now exposes an optimizedForCode capability on text models. Docs have been updated.

• The model list endpoint now exposes a supportsVision capability on text models. Docs have been updated.

• API Key expiration dates are now returned on the Rate Limit endpoint. Docs have been updated.

• The model list endpoint now exposes all image constraints that are part of the schema validation for image generation. Docs have been updated.

• Postman Authorization helpers have been configured for Venice’s collections. This should help provide instructions for new users on how to generate their API Keys.

• Fixed a bug in the image generation API that was causing content types of binary images to be returned as image/undefined. Added a test case to avoid regression.

• Fixed a bug that was preventing models that had the supportsResponseSchema capability, but not supportsToolCalling from properly processing response_format schema inputs.

• Fixed a bug where Brotli compression was not successfully being passed back to the API caller. The postman example has been updated and a test case has been added.

• The Postman test suite has been completely overhauled and optimized and integrated as part of Venice’s broader CI pipeline.

Venice Voice Processes over 1 Million Sentences in its first 24 hours…

Over the last 24 hours, Venice Voice has processed more than 1 million sentences. We’re thrilled to see the interest in this offering and look forward to including its capabilities via the API in the coming weeks.

Qwen QwQ 32B available for Pro and API users

Today, we enabled Qwen QwQ 32B for all Pro users and API users. Per the Qwen docs, QwQ is the reasoning model of the Qwen series. Compared with conventional instruction-tuned models, QwQ, which is capable of thinking and reasoning, can achieve significantly enhanced performance in downstream tasks, especially hard problems. QwQ-32B is the medium-sized reasoning model, which is capable of achieving competitive performance against state-of-the-art reasoning models, e.g., DeepSeek-R1, o1-mini.

Venice’s QwQ implementation is web enabled and supports structured responses.

This model replaced Deepseek R1 70B as our medium weight reasoning model.

App

• Per this Featurebase request, generated images now contain EXIF metadata including the generation settings used to create the image. This metadata can be disabled in the Advanced Image Settings.

• Made numerous updates to Markdown rendering in chat to fix issues with ordered lists and code blocks.

• Permit WebP images to be uploaded by dragging and dropping them into the chat input.

• Optimized Venice Voice sentence chunking to ensure invalid characters aren’t sent to the voice model, and to reduce network overheard for longer conversations.

API

• Using the new format parameter, the API can now generate WebP images. Images currently default to PNG to maintain compatibility with prior generation. This will change in the future, so please add a specific format to your API calls if the response format is important. The docs have been updated.

• EXIF Image Metadata can be added or removed from generated images using the new embed_exif_metadata parameter. The docs have been updated.

• Reasoning models now expose a supportsReasoning capability in the models list endpoint. The docs have been updated.

• Fixed a bug where the Rate Limit API would not show default explorer tier limits.

• Removed the admin key requirement on the Rate Limits and Balances endpoint and the Rate Limit Logs endpoint.

• Remedied Swagger validation issues in our published swagger docs and added a step to our CI pipeline to ensure future validation.

Announcing Venice Voice Preview for Pro Users

We’re thrilled to release Venice Voice, our number one most requested feature from Featurebase. Powered by the Kokoro model, Venice Voice brings Venice to life by reading responses in a wide variety of voices. Please note, this feature is in preview and we welcome your feedback and bug reports while we continue to work towards the final production release.

This feature can be enabled both directly on a message by clicking the speaker:

Voice and speed of Venice Voice can be controlled from the text settings:

Venice Voice is now also supported on Characters. Character creators can pick a voice and speed for their character, and Venice Pro users with Venice Voice enabled will be able to have the character responses read back:

API

• Changed a few default settings on image generation and updated the docs to accordingly.

• Added a way to delete historical files from the context.

App

• Fixed a bug where stopping a message on a reasoning model could potentially discard the thinking process content. Thinking process will now be maintained when a message is stopped.

• Fixed a bug where ordered lists may render oddly when produced from LLMs.

Consumption Limits for API Keys

Venice users can now define consumption limits per epoch in both USD and VCU on API keys in both the web UI and via the API. This allow users wishing to share their VCU with others users the ability to parcel out specific allocations on a key by key basis. The API docs have been updated here and the Postman Create API Key example has been updated.

Multi-Image Variants

Venice Pro users can now generate multiple image variants in a single request. Configurable within the app’s “Image Settings” panel, you can select between 1 and 4 variants. Once generated, clicking on a variant will give you a zoomed in view and you’ll find action buttons in the bottom right corner.

WebP Image Support

For image generation, Venice switched the default response type to WebP. WebP is an image format that enables smaller, faster-loading images and given Venice stores all generated images in the browser, smaller images will make the app experience faster.

For users who wish to continue to use the PNG format, you can change the setting in the Image Settings window:

API

• Added a Rate Limit Log endpoint. This will return the last 50 rate limits an account has hit and is designed to help diagnose rate limit violations. Docs have been updated and a Postman example has been created here.

App

• Removed “High Res” Image Settings — this was causing inconsistencies in the quality of image generation and with the new 2x upscale option that was added, Venice users can get better quality images using the standard sizes and then upscaling.

• Fixed a bug where regenerating a failed message could potentially change the model.

App

• When the account level “Safe Venice” mode is enabled, Uncensored image models will now be redacted from the model selector.

• Provide more human interpretable errors when uploading files to the chat input.

Characters

• Character chat now has a web enabled toggle as part of the input.

API

• Added support for response_format in our API. The docs have been updated and an example can be found here in Postman.

• Added supportsResponseSchema capability in the API models list. The docs have been updated.

• Added a default_reasoning API trait assigned to Deepseek 671B. See this endpoint for a full list of all API traits.

• Exposed all API headers to be visible in the browser via Access-Control-Expose-Headers

• Resolved issues with x-venice- headers in API responses not consistently returning rate limits and VCU / USD balances.

App

• Filtered out uncensored image models from the available image model list when Venice Safe Mode is enabled for image generation.

• Hide chat settings on character pages unless toggled on.

API

• Added documentation on the /characters endpoint that provides details on available public characters.

• Resolved issue with in-painting where invalid images would throw 500 errors. Invalid images will now throw 400 Bad Request errors with a description in the body of the response.