DeepSeek Harness got a lot of people excited, but I used it once and didn't keep going; Qwen3.8 27B's capabilities have taken off, but I haven't deployed it locally either. Weighing the urge for new things against the need for stability, OpenCode + Paseo + OpenViking is already smooth enough. Anything new that doesn't offer real, controllable, reliable help is just noise.

When using QwQ-32B, the MLX version throws this error during chat: “Error rendering prompt with jinja template: Error: Parser Error: Expected closing statement token. OpenSquareBracket !== CloseStatement.”, which you need to fix by modifying the jinja template. The gguf version doesn't have this problem. After the fix it ran successfully, as shown in the image.

Trae's China version now supports DeepSeek, and it outperforms Claude3.5-Sonnet in some scenarios. QwQ-32B's performance is close to R1, making it a good fit for local deployment. If Manus goes mainstream, it will upend existing technology, and future AI will need cloud resources to support multimodal operations and agent collaboration.

It's Chinese New Year, and everyone is talking about DeepSeek. Ruirui wants to play "Who's the Spy" but can never round up enough people, so I built an AI-powered "Who's the Spy" game bot that calls an LLM to generate character descriptions and voting strategies — but I ran into problems like unstable model output and formatting errors. After tweaking the model and the prompts over and over, I finally got the basic functionality working, and the whole process was full of funny mistakes and challenges.

I'd like to recommend AnythingLLM. Combined with LM Studio, it can run local large models as a private AI assistant. I'll share my experience trying it out for document processing and studying traditional Chinese medicine. The results are only so-so for now, but the journey from zero to one is well worth it.