Basalt 0.4.0 introduces multi-model and router support, starting with DeepSeek V4.1 Flash, alongside video capabilities and file uploads in chats. The release adds Qwen 3.8 27B as its first dense model, achieving up to 770 tokens per second on an RTX 5090 while maintaining 92.8% top-1 accuracy versus the full BF16 model. This update lays groundwork for the upcoming Qwen 4 integration.
Read original
reddit/r/LocalLLaMA