jan
jan copied to clipboard
Jan is an open source alternative to ChatGPT that runs 100% offline on your computer. Multiple engine support (llama.cpp, TensorRT-LLM)
Regenerating answers leads to strange output ![image](https://github.com/janhq/jan/assets/6628064/dccf78a1-e507-4a24-b350-8f6294750172) ![image](https://github.com/janhq/jan/assets/6628064/73268be5-fa5e-4bd6-94ac-8ed4667dc590) Using Mistral 7B Instruct v0.2 Q5_K_M with chat prompt `[INST] {prompt} [/INST]` running on Vulkan acceleration.
**Problem** DeepInfra supports many open-source models via their API. **Success Criteria** User can use models from DeepInfra with ease. **Additional context** Link: https://deepinfra.com/models
**Describe the bug** When starting a new thread and interacting with the model, I can see after the first prompt the thread-title changed. However, after closing and reopening Jan, the...
## **1. Release scope:** #### :truck: Main features: * epic: Import model via Huggingface URL #1740 * IQ quants support #2631 * feat: Support Command R+ #2678 * Compatible with...
**Specs** https://www.notion.so/jan-ai/MVP-Whitelabel-support-via-a-Config-file-or-Manifest-2124dee2055f4954a7495a08de110980 **Success Criteria** We have a config file for fast deploy app **Additional context** TBD
**Bug Description** By default Jan will choose CPU mode. My local AI models are marked as "Recommended". When enabling Experimental Mode and Vulkan support, all local models are marked as...
## Motivation Model from API providers should be more clarifier ## Specs https://www.notion.so/jan-ai/Refactoring-of-Remote-API-Extensions-Mistral-Cohere-etc-8638a1fa26ca48f3b57822157c11152a?pvs=4 ### In-scope - Cohere extension https://github.com/janhq/jan/issues/2686 - Anthropic extensions https://github.com/janhq/jan/issues/2777 - Deepinfra extensions https://github.com/janhq/jan/issues/2717 - Openrouter https://github.com/janhq/jan/issues/2685
**Describe the bug** Context length and NGL are two critical settings that can help users overcome issues with loading heavy models on their machines, such as hogging or OOM errors....
**Problem** Openrouter is a service that helps users to use the best model for each query E.g. Easy query like hello → use Mistral, Hard → use command R+, etc...
**Problem** Command R+ is the best open-source model that can surpass GPT-4 in some benchmarks. However, the model is 104B so normal local PC can't run this model. We should...