Trae
Three days ago, Trae released its China version, hooked up to DeepSeek. In some cases R1 seems more dependable than Claude 3.5 Sonnet.
When Claude 3.5 Sonnet helps me port big chunks of my old code, it sometimes goes off-script and slips bugs in instead — and then I end up burning a ton of time debugging it myself.
It’s great at writing simple web stuff; going from 0 to 1 with it is genuinely impressive. But on anything more complex it keeps getting it wrong over and over, like an intern. I honestly don’t get how the people out there hyping Claude so hard are actually using it.
QwQ-32B
Today QwQ-32B became installable, so I jumped on it right away.
Word is it’s as strong as the 671B R1, which would make it a great fit for local deployment.
Manus
Well damn — if this really becomes as widely accessible as DeepSeek and Qwen, then all the setup I’m learning right now will be obsolete.
Official site:
It probably flexibly folds in every trick that’s actually usable in practice right now, and pushes each one to its limit.
Or a product like Manus gets used in enterprise-grade ways. Local super-individuals just keep writing a bit of code themselves and cobbling together some local AI function calling, and that’s enough.
So future enterprise-grade AI platforms are far from done at multi-dimensional tables for batch-scheduling parallel tasks, workflows with multiple AIs and multiple modalities working together, and knowledge base chunking and extraction — they’ll also have to provide cloud PCs for multi-modal agents to operate. A general AI assistant needs a brain and needs hands, and it’s exactly these cloud resources that a standalone version can’t match.