Changelog

App

• Updated image rendering performance on conversations with long histories.

• Updated the rate limit display to populate on first page load.

• Simple Mode - Upgraded the model router to our latest variant which has superior routing and added support for additional languages.

• Simple Mode - Improved the loading state when a message is sent.

• Simple Mode - Persist the status of simple mode on a user’s logged in account between devices.

• Improved performance of image upscaler / enhancer.

• Fixed an issue that would prevent PDF parsing of documents with custom fonts from failing.

Characters

• Added a toggle to enable / disable reasoning mode on supported models.

API

• Released the embeddings endpoint for all users.

• Added the Venice “Name” to the Model List endpoint. API docs have been updated.

Models

• Updated FLUX Custom to utilize an improved fine tune model.

• Retired Fluently and Stable Diffusion 3.5.

App

• Increased context size on Venice Large from 32k to 128k tokens.

• Changed the default image model from Venice SD3.5 to HiDream.

• Changed image generation to show each image as it’s generated instead of waiting for them all to complete to display. This should result in much faster time to first image.

• Simplified the duplicated “Safe Venice” settings into a single “Mature Filter” setting that optionally supports setting a PIN.

• Added in-line citations to search results.

• Improved image rendering performance to reduce memory pressure on browsers for conversations with long image histories.

• Add a setting to the “App Settings” that allows users to mask their personal information in the user interface:

  

• Fixed a bug that was showing a success message on shared encrypted links when the sharing was unsuccessful.

• Fixed a bug where images showing up from variant generation did not appear in the proper order when viewed in the light box.

• Updated the chat input prompt text to indicate users can ask questions about an image when one is attached in chat mode.

• Increased anonymous user image generation limits to 16 per day.

• Substantial improvements to model settings handling focused on ensuring optimal settings are selected from model to model. 

• Fixed a bug that could cause Venice Small and Venice Reasoning to hallucinate references when no web search was conducted.

• Fixed a bug where double scroll bars could show up on chat settings.

• Fixed a bug that would cause some closing reasoning (</think>) tags to be ignored, causing improper rendering in the UI.

API 

• Increased context size on Venice Large from 32k to 128k tokens.

• Fix for thinking parsing to ensure the values set in venice_parameters are properly interpreted and displayed in the response. Ensures specific think messages are properly stripped in non streaming mode.

Last week, our engineering team focused heavily on platform stability and reliability along with preparing a number of new features to release to our beta test group. We look forward to sharing those out over the coming week.

App

• Images will now render in the UI as they’ve completed vs. waiting for all four to complete for display.

• Added (i) info buttons to text chats to permit a view of the details of the message (like temperature and top-p).

• Added sign-up link in the left navigation menu when signed out.

• Added UI elements that display upscale/enhance images in the chat history. Upscales > 1x will still be automatically downloaded.

• Added a UI timer to the upscaler message UI to provide some sense of how long the process has been running.

• Clarified wording on error messages related to capacity on upscale / enhanced images.

• Fixed a rendering bug related to escaped dollar signs on code blocks in chat.

• Fixed model specific settings not appropriately applying when selecting “Use these settings”

• Fixed an issue with default settings not resetting across all models as expected.

• Fixed a bug related to the “download all images” button missing on new chats.

• Deprecated the code specific section of the web app to streamline the web UI.

API

• Updated upscale/enhance API to provide more clear error messages when images are outside the bounds of supported sizes.

• Added support for top_logprobs parameter as part of the chat completions API.

• Optimized the performance of the API keys list endpoint used in the app UI and available in the API particularly benefiting users with substantial API traffic.

Given the short week, the team’s engineering efforts this week was primarily focused behind the scenes on reducing model error rates, and on a set of new features that are coming soon.

App

• Added a hover action state on images.

• Fixed a bug preventing the mobile share screen from appearing when with sharing an image on mobile devices.

• Fixed a bug that prevented horizontal scrolling when viewing complex LaTeX formulas wider than the viewport.

• Fixed a bug that would cause images to stop displaying after viewing the (i) info details.

• Disabled negative prompt on HiDream.

API

• Added support for multiple image processing to Qwen VL.

Characters

• Added new Character setting to hide model reasoning / thinking output for relevant models on character conversations.

• Updated “Reasoning” text to reference “Thinking” for character conversations.

It’s been a busy week for our team with 3 major releases, and a slew of updates and bug fixes for you:

• Simple Mode for Pro Users

• Venice… Enhance any Image

• New Image Model: HiDream

  ---

Pro User? Help us test Simple Mode

We're about to roll out the biggest upgrade to the Venice's user experience since our launch: Simple Mode.

Simple Mode provides a streamlined experience by automatically routing your prompts, eliminating the need to manually select a chat type or model (if you ask for an image, it will just know it, without any conception of "Image Mode").

We’ve also added the concept of "image context", which allows you to build images through subsequent prompts. We believe for new users, this will dramatically simplify the user experience, making Venice feel and behave more like other generative AI platforms.

How can you help?

1. Go to your App settings and toggle on Simple Mode. You will see that the model dropdown, and mode selector, will disappear.

2. Chat with Venice and try different prompts in natural language. "Tell me about France" or "Create image of France" etc. Text, web search, image generation, all of it should work automatically.

3. Send your experiences and feedback to support@venice.ai. Please include screenshots and videos, or share your chat for things that don't seem to work as expected.

We appreciate you taking the time and look forward to hearing your feedback.

Venice… Enhance Image

Pro users can now upscale and enhance any image. Click on the paperclip or drag and drop an image into the image generation chat.

Upscale & Enhance system represents a significant leap beyond traditional pixel-by-pixel methods:

• Use Upscale for accurate enlargements: Use the Upscale feature when you want to maintain the exact style and composition of your image while increasing resolution.

• Use Enhance for creative enhancements: Use the Enhance feature when you want to add details and refinements beyond what's in the original image.

HiDream is now available on Venice

Read more about our latest model in our announcement tweet.

Key Features Prompt adherence - Achieves industry-leading scores on GenEval and DPG benchmarks, outperforming all other open-source models.

Text placement - Effective and accurate text generation and placement, making it useful for graphic design purposes.

Commercial-friendly - Generated images can be freely used for personal projects, scientific research, and other applications.

App

• Added an image generation loading state to make it more clear that the UI is waiting on inference.

• Added a setting to enable display of message timestamps in the App Settings - Featurebase

• Fix crash when editing a chat message and using Space key. - Featurebase

• Updated the error when documents too large to be processed to be human readable.

• Refactored context filtering logic to prevent Your message is too long for the selected model errors when dealing with chat conversations with large histories.

• Ensure that white space submitted in system prompts sent from the app are preserved all the way to the LLM.

• Improved error messages when there are issues communicating with Venice’s inference servers.

• Update the image generation prompt enhancer to ensure messages are constrained to fit within the image prompt limits.

API

• Fixed a regression in streaming responses when invoking tool calling. Added test coverage to avoid future regressions.

• Enabled support for response_format to 405B and Deepseek. Ensure thinking models don't inject <think> tags into the response when response format is enabled.

• Return 504 timeouts when non streaming requests timeout. Recommend use of streaming responses when inference is expected to run for longer durations.

• Remove default `max_tokens` on API calls, permitting inference to use the full context of the LLM unless the user specifies a specific max_tokens

Characters

• Introduced a character avatar popover on larger screens, displaying detailed character information and a quick chat option.

• Added display of character creation date and creator information on public character profiles.

• Introduced clearer visual grouping and dividers in character menus.

• Added a "View & Rate Character" menu item for improved access to rating features

Token

• Updated the token dashboard to fix a few UX issues when used via mobile wallet.

The big release over the last week was the launch of Venice Search V2 - a complete overhaul on how our search function operates. This was implemented for both our App and API users. Venice search is now:

Smarter
Now uses AI to generate search queries based on chat context rather than directly searching input text. This results in more contextually relevant information being injected to the conversation, and better overall responses.

Cleaner
Only displays sources actually referenced in the response, using superscripts. These reference the citations provided below the search.

Broader
We inject a greater number of results with additional information per result into the context.

App

• Released Venice Search V2.

• When switching models, the Top-P and Temperature settings will now automatically default to the optimal setting for that specific model. Additionally, a UI element was added to show what that default for the model is. This should remedy issues with temperatures changing as users move through models resulting in potential gibberish in responses.

• Adjust the “image prompt enhancer” to keep its responses below the character limit for image generation.

• Add a link to the hugging face model card from within the Image Detail view.

• Add a w/ web search banner to responses that have included web search.

• When using shorten or elaborate, the current selected model will be used for the response, vs. the model that the original message was generated from.

• Using the space bar will now trigger the “accept” button within confirmation screens.

API

• Released Venice Search V2.

• Added support for purchase of API credits with Crypto via Coinbase Commerce.

• Add support for strip_thinking_response for reasoning models. This will suppress the <think></think> blocks server side, preventing them from reaching the client. Works in tandem with /no_think on the Qwen3 models. API docs have been updated for the parameter, and the model feature suffix docs have also been updated. Satisfies this Featurebase.

• Add support for disable_thinking for reasoning models. This will add /no_think in the background, and enable strip_thinking_response - API docs have been updated and the model feature suffix docs have been updated.

• Add support for enable_web_citations - This will instruct the LLM to reference the citations it used generating its responses when Web Search is enabled. API docs have been updated and the model feature suffix docs have been updated.

• Remove 4x option and show "max" in its place. This will leverage the above change on the API to allow images that can't 4x upscale to be uploaded. This will still block images that are > 4096 x 4096 since the scale can't be less than 1.

• When upscaling, if scale is set to 4, dynamically reset it so that the maximum final output size is always less than the max pixel size of our upscaler.

• Added a model compatibility mapper for gpt-4.1 to map to Venice Large / Qwen 3 235B.

• API Key Creation is now rate limited to 20 new keys per minute with a total of 500 keys per user.

Characters

• Added a limit to character names to prevent issues within the UI.

• Fixed up character display for characters with excessive display information that was previously breaking the page layout.

• When using the auto-generate character feature, a confirmation box will be presented first to avoid overwriting existing details on accident.

It’s been a big week of updates with many additional updates and enhancements coming over the next few weeks. Please check out our beta group on Discord if you’d like to participate in early testing of our new features.

Model Updates

Thank you to the community for all the helpful feedback after our new model paradigm launched. We will always refine our offering and your opinions are immensely helpful. We've decided to make a couple model changes.

• Llama4 Maverick (aka Venice Large) is being retired May 12th
  Zuck failed us on this one and it needs to go. It has been replaced with the new Qwen3 235B as our Venice Large model. The Venice beta users have been enjoying it, almost universally prefering it to Maverick.
  

• Llama 3.2 3B (aka Venice Small) is also being retired.
  This model had a good run, but a plainly superior option exists now. The new Qwen3 4B will replace it as our Venice Small model. Beta users also very positive on this one. Both of the retiring models will remain in the app+api for 2 weeks under their own names. Maverick will then be taken out to pasture, removed from both app+api. Llama 3.2 3B will leave the app, and remain in API for some time.
  

• Deepseek retirement has been postponed to May 30th
  Additionally, we’ve heard your feedback RE: Deepseek’s retirement and we’re thinking through options. The retirement for Deepseek has now been moved to May 30th and we’ll provide another update before then.
  

• Inpainting Deprecation
  We are re-engineering Venice’s in-painting feature set to better serve the use cases we’ve now seen from our users. We are going to deprecate the current version from the app and the API next Monday while we work on the new release.
  
  In the interim, we encourage users to experiment with Venice’s “Enhance image” feature which can create neat re-creations of images.

App

• We've released an update that should alleviate grammatical errors and missing characters from longer conversations, most notably on 405B. If you continue to see those issues, please use the report conversation feature. Thank you for the existing reports -- they were very helpful in tracking down the issue.

• Updated the Report Conversation feature to allow for self reported categorization of the issues. This helps our team identify trends and issues with models faster.

• Added a Reasoning toggle for reasoning models that support enabling or disabling thinking responses.

• Added a warning within the chat input for users who have increased temperature into bounds known to create gibberish / garbage responses.

• Updated the Venice system prompt to reduce likelihood of Venice referencing details about itself in responses unless prompted about Venice.

• Streamlined the share chat functionality to immediately copy the share URL to the clipboard vs. requiring a second click.

• Updated the UI to disable the upscale / enhance button if both upscale was turned off and enhance was disabled.

• Updated the UI to only copy the user’s prompt when copying prompt + images messages.

• Updated the UI to view image options when viewing image variants in grid format.

• Fixed a bug where non-pro users were unable to upload documents or images for analysis.

• Fixed a bug when editing messages containing code blocks that would result in certain characters being improperly escaped.

• Ensure full EXIF / ICC profile is maintained when using the upscale / enhance feature. Fixes this Featurebase request which had two [1] [2] new reports.

API

• Security Notice - Fixed a bug reported via our bug bounty program that permitted API keys marked as inference only be able to manipulate the API key admin endpoint. This would have permitted these inference only keys to add, or remove other API keys. Please validate active API keys created between April 22nd, 2025 and May 7th, 2025 to ensure their validity.

• Explorer Tier Deprecation - As Venice continues its growth, we’re seeing our API usage reaching all-time highs. Following our announcement last month, we have changed our Pro account API access.
  
  Previously, Pro users had unlimited access to our Explorer Tier API with lower rate limits. We have now deprecated the Explorer Tier, and all new Pro subscribers will automatically receive a one-time $10 API credit when they upgrade to Pro –double the credit amount compared to competitors.
  
  This credit provides substantial capacity for testing and small applications, with seamless pathways to scale via VVV staking or direct USD payments for larger implementations. This change reflects our API’s maturation from its beta to the enterprise-ready service that developers are increasingly building on the API.

• Ensure full EXIF / ICC profile is maintained when using the upscale / enhance feature. Fixes this Featurebase request which had two [1] [2] new reports.

• Add support for OpenAI Embedding names to the Embeddings API via Model Compatibility Mapper.

• We've released an update that should alleviate grammatical errors and missing characters from longer conversations, most notably on 405B. If you continue to see those issues, please use the report conversation feature. Thank you for the existing reports -- they were very helpful in tracking down the issue.

• Add support for JSON payload to Upscale / Enhance API - API docs are updated - Postman example.

• Fixed a bug that caused the created field on the OpenAI compatible image generations API to ensure it's coming back as an int and not a float.

• Fixed a bug causing model_feature_suffix features from properly updating their respective flags. Added additional test coverage to ensure this avoids a regression.

Token

• Updated the token dashboard coloring.

• Redirect identifiable mobile wallets to the token dashboard when accessing https://venice.ai and hide the PWA installation modal.

Characters

• Updated Character UI with imports and rating stats on the primary character cards.

• Added a UI feature to show which source character a user’s character was cloned from.

New Model Paradigm

Venice simplified its model selection with a curated list of LLMs, categorized into five distinct models: Venice Uncensored, Venice Reasoning, Venice Small, Venice Medium, and Venice Large. You can find more details in our blog post here.

The new models include the Dolphin Mistral 24B Venice Edition, Venice's most uncensored model ever, and Llama 4 Maverick, a vision-enabled model with a 256K token context window. Several legacy models, including DeepSeek R1, Llama 3.3 70B, and Dolphin 72B, will be retired from the chat interface by May 30. The changes aim to reduce model redundancy, improve user experience, and increase infrastructure scalability.

All current models remain available through the Venice API.

App

• Implemented a substantial revision to search behavior to ensure search results are more effectively integrated into the context.

• Added support for “Enhance Only” mode via the app. This permits the endpoint to be used solely for enhance without changing the output resolution of the image:

• Added a prompt for users to permit the browser to persist local storage when their browser storage is becoming full.

• Fixed a scrolling bug for users with character chats per this Featurebase.

• Added some guidance to the app suggesting using descriptive prompts or the enhance prompt feature when using the Venice SD35 image model.

• Added in-app guidance when the Temperature setting has been set very high to indicate the LLM may return Gibberish.

• Added a subscription renewal flow for Crypto user’s who wish to renew their subscription.

• Fixed a bug where upscale / enhance requests could return blank / black images.

• Adjusted the pre-processing for in-painting to increase reliability of generation.

• Fixed a bug where the Input History Navigation setting in App settings was not properly controlling the feature behavior per this Featurebase.

Characters

• Improved character search UI.

• Update UI to permit free and anonymous Venice users to see the characters detail modal.

API

• Added pricing information to the /models endpoint per this request from Featurebase. API docs have been updated.

• Increased Token per Minute (TPM) rate limits on medium and large models given Maverick can produce a large number of tokens quickly. API docs have been updated.

• Added support for a 1x scale parameter to the upscale / enhancement API endpoint. This permits the endpoint to be used solely for enhance without changing the output resolution of the image. Solves this Featurebase. API docs have been updated.

• Added a new API route to export billing usage data. API docs have been updated.

• Added support for the logprobs parameter on our /chat/completions API. API docs have been updated.

• Added a UI to the API settings page to export billing history from the UI.

• Added support for fractional scale parameters to the upscale / enhancement API endpoint.

• Updated the API to require application/json headers on JSON related endpoints.

• Return additional detail in error message if a model can not be found to assist user’s in debugging the issue.

• Added support for Tools / Function Calling to Maverick.

• Launched an /embeddings endpoint in beta for Venice beta testers. API docs have been updated.