v0.1.7

What's Changed

• Fixed installer to clean install the application, avoiding weird bugs
• Fixed application and server crashed on large prompt
• Added control of maximum number of tokens to be processed on each iteration frame
• Fixed chat name can't contain some symbols
• Allowing to rename duplicate chat name
• Fixed pasted long text on system prompt crashed the application
• Acrylic background
• Refactor AI model config
• Added section of downloaded model
• Sort model list based on alphabet
• Added search in model manager model
• Gemma 3 support!

Full Changelog: v0.1.6...v0.1.7

v0.1.6

• Introducing Kolosal AI Server, easily managed server within kolosal AI's application.
• Added phi 4 and phi 4 mini models.
• Added continous batching mechanism for decoding
• Added kv cache management mechanism for batch decoding
• Added model loading settings within server tabs
• Added tab management system
• Added automatic title generation for each chat history

v0.1.5

What's Changed

• Context shifting with StreamingLLM (https://arxiv.org/abs/2309.17453) to unlimited generation
• Llimit max context to 4096 to make it more memory efficient and faster
• Added stop generation
• Added regenerate button
• Redesign progress bar
• Model loading handled asyncrhonously
• Added unload model button
• Huge refactor
• Fix code block rendering glitches
• Setting max new tokens to 0 will result in unlimited generation with context shifting
• Fix application crash when delete a chat

Full Changelog: v0.1.4.1...v0.1.5

v0.1.4

• added deepseek r1 support
• added markdown rendering
• added tps stat
• added cancel download button
• added delete model button
• fixed model duplication issue
• fix engine memory leak
• added thinking UI
• add automation to detect number of thread to use
• fix last selected model issue
• add fallback on loading model failed

v0.1.3

New feature

• Added persistence KV Cache method: Persistence KV Cache allow Kolosal to have model kv cache state saved for each chat history, making processing previous chat to be instant.

Bug fixing

• Fix model parameter is not correctly passed to the model
• Fix deleting a chat, crashed the application
• Fix switching model resulting the model to generate in different chat
• Fix does not detect AMD GPU
• Fix EOS not detected on a finetuned model with chatml format
• Fix force Close in chat feature
• Fix performance issue on GPU

New model

• qwen 2.5 code 0.5b - 14b
• qwen 2.5 14b

v0.1.2

What's Changed

• Fix GPU support, now can detect automatically nvidia/amd gpu in your device and select it
• Added clear chat and delete chat buttons
• Fix application shortcut (removed fn + left arrow key short cut to open kolosal)
• Added Qwen2.5 models 0.5 - 7B

Full Changelog: v0.1.0...v0.1.2

v0.1.1

What's Changed

• Added Windows Installer
• Added Sahabat AI Llama 3 8B
• Added Sahabat AI Gemma 2 9B
• Added Gemma 2 2B
• Added Gemma 2 9B
• Added Llama 3.1 8B
• Added 8bit quantization support
• Update quantization selection UI to be radio button

Full Changelog: v0.1...v0.1.1

v0.1.0

Kolosal AI 0.1 marks the very first release of our groundbreaking solution for on-device Large Language Model (LLM) inference. Engineered to run smoothly on both CPUs and a variety of GPUs, Kolosal AI brings powerful AI capabilities to Windows 64-bit systems without relying on external servers or cloud dependencies.

Key Features and Highlights:

• On-Device Inference: Harness the power of advanced LLMs locally on Windows 64-bit machines, preserving data privacy and reducing latency.
• Broad Hardware Support: Optimize performance on most common CPUs and GPUs, making Kolosal AI accessible and efficient for a wide range of hardware configurations.
• Easy Integration: Seamlessly incorporate Kolosal AI’s inference engine into existing applications or workflows with straightforward setup and minimal dependencies.
• Low Latency & High Throughput: Experience near real-time responses through efficient model optimization and hardware utilization.

With Kolosal AI 0.1, developers and enthusiasts alike can begin exploring the possibilities of large-scale AI right from their desktop—no cloud required. This release is just the start of our mission to empower everyone with the latest advancements in artificial intelligence, all within a user-friendly and hardware-agnostic platform.