Latest Updates & Improvements to Faraday
- Small fixes & improvements
- Optimized Character Hub homepage queries for faster loading
- Added pruning for unnecessary prompt template tokens in Character example dialogue
- Fixed default User persona image missing on newly created Characters
- New Cloud model
- Added Llama 3 Lumimaid 8B to Standard and Pro plans
0.18.17 - 0.18.19 Internal
- Impersonate feature
- Added ability for the model to generate user responses
- The Impersonate button is located to the right of the Continue button
- Improvements to "Experimental" backend on Desktop
- Improved GPU detection (note: models may load slowly the first time after this update, but subsequent loads will be fast)
- Fixed Llama 3 response quality issues related to the tokenizer
- Increased token rate by 5-10% on Apple metal and CUDA
- Fixed tokenizer issues affecting Command-R, Qwen2, DBRX, and other base model architectures
- Added flash attention optimization (does not apply to Vulkan)
- Fixed gibberish responses when using Vulkan GPU acceleration
- Bug fixes & improvements
- Fixed issue preventing chat deletion in the header dropdown
- Improved Llama 3 prompt templating
- Removed hardcoded separator message between example dialogue and chat history
0.18.13 - 0.18.15 Internal
- Llama 3 Fixes
- Fixed low quality outputs when running Llama 3 models running at >4096 context
- Better ROPE calculation for Llama 3 GGUFs downloaded outside of Faraday
- Cloud Update
- PRO: Added Llama 3 Ultra Instruct 8B at 16,384 tokens of context
- Deprecated Llama3 Solana 8B
- Fixes & Improvements
- Removed word limit on User Persona templates
- Removed “initializing model” loading message on chat page
- Fixed Character settings updates while Tethering
- Fixed chat log imports from external apps (Ooba, CAI tools)
- Fixed character imports from Faraday PNGs
- Fixed character settings page reverting input fields after refocusing the window
- Fixed issue pulling in image updates from Hub Characters
- Added ability to sort by recently updated characters on Hub
- Updates to “Experimental” backend
- Fixed low quality outputs when using Llama 3 GGUFs (all Llama 3 models will need to be re-downloaded from the Model Manager)
- Added support for base model architectures including Phi, SEA-LION, Command-R, and more
0.18.10 Internal
- Reusable User Personas
- On the General Settings page, you can now create reusable personas that can be applied to any character
- Each persona includes a name, description, and image
- Llama 3 Inference Fixes
- Fixed repeating issues for most models
- Fixed issue where </s> would output at the end of every generation
- Added better ROPE calculation for Llama 3 models at long context lengths
- Better GPU detection on Experimental backend
- Cloud Update - Llama 3 Models
- PRO: Added Llama3 Solana 8B at 16,384 tokens of context
- STANDARD: Added Llama 3 Soliloquy 8B at 4,096 tokens of context
- Bug fixes & improvements
- Fixed edit message button always showing on mobile web
- Minor UI fixes on chat page & character settings page
0.18.5 - 0.18.8 Internal
- Hotfix
- Fix issue where characters would forget recent context
- Fix cloud streaming on desktop
0.18.2 - 0.18.3 Internal
- Support for Llama 3 base model
- If using a custom model, we recommend that you use the "Llama3" prompt template in the Character settings
- New "Experimental" Backend in Advanced Settings
- Support for all IQ quants
- Performance improvements for CUDA, Vulcan, and Metal
- Up to 40% faster prompt evaluation speeds for CPU
- Faster inference for chats that include grammars
- Cloud Infrastructure Improvements
- Significantly reduced wait times for all PRO plan models
- Faster token rates for Midnight-Rose 70B
- Faster token rates for all STANDARD models
- Fix for occasional “socket” errors after sending a message
- Bug fixes & Improvements
- Better mobile chat UI
- Increased max length for author’s comment on Hub
- Fix overflow for input fields on the Character creation page
- Fixed infinite loading on the Advanced Settings for some Windows devices
- Confirmation warning before irreversible undo operations
- Delete cached images after deleting a Character
- Fixes "maximum call stack exceeded" error when exporting a Character card
0.18.0 Internal
- Multiple Character images
- Support for adding up to 10 images to a Character
- Ability to scroll through images in the chat page sidebar, and on the Character Hub
- More TTS voices
- Bug fixes and improvements
- Enabled TTS voices in the browser
- Add alphabetical sorting to local homepage
- Increase default model context to 4096
- Fix issue where clicking undo would focus the button
- Fix to prevent overlapping TTS voices
- Fixed issue where GPU settings page would not load
- Fixed issue where images with the same name would overwrite each other
0.17.5 - 0.17.6 Internal
- Bug fixes and improvements
- Improve model context management for new example dialogue pruning feature
- Fix for occasional empty responses from the model
- Fix "invalid access to closed websocket" error
- Fix for missing and duplicated regenerations
0.17.3 Internal
- Hotfix
- Fix for issue preventing app startup when offline
- Fix for UI issue preventing menu from showing on homepage character cards