MateEngine 2.2.0
Huge AI Update
Switched from the local LLM "Qwen 3 4B" to "Llama 3.2 3B" by Meta Platforms Inc. We found that Llama 3.2 3B fits much better for casual conversation, and the previously known "Loop" text issue has been fully resolved.
Conversations now feel more natural and coherent.
The Chat Bot no longer repeats or rephrases the same answer.
Markdown rendering has been improved for better formatting and readability.
The new model is more lightweight and performs faster on mid-range GPUs.
VRAM usage has been reduced from approximately 6.5 GB to 4.5 GB on average.
Added a Context Length setting that lets you choose between 2K, 4K, 8K, 16K, and 32K. Keep in mind that higher context sizes require more VRAM and power. The default value is 4K, which runs well on low- and mid-range GPUs. For the best performance, we recommend 8K or 16K.
Overall chat streaming speed has been greatly improved.
The total build size of MateEngine is now 500 MB smaller.
Built with Llama 3.2 — © Meta Platforms Inc.
Not affiliated with or endorsed by Meta. Not an official Meta product.