Ox Alpha (GLM-5.3-Flash) just got exposed.
Same prompt. Same harness. Same effort. DeepSeek-v4-Flash: 10 min, $0.02 GLM-5.3-Flash: 73 min, $0.16 GLM made the richer scene. But at 7× the time and 8× the cost… is it really better?
Qwen turned me on again. 🔥 The 176B level stuff can’t be viewed only with the computer room card.
Fortunately, Liangzi installed eyes on DeepSeek last week. Otherwise, when Qwen comes out, the scene today will be somewhat embarrassing. 😂 DGX Spark can run. Dual 4090 96G will also work. Throw the 51B N-gram Embedding into the CPU memory and put the rest into the video memory.
Opus 5 - > Sol 5.6 - > Grok 4.6 - > deepseek v4 pro
All in high thinking, simple prompt, without skills. Tried with gemini 3.7 with api but impossible to use. No matter how many crazy things Opus 5 talks about, it was the clear winner.
The GitHub Wanxing Team (MemOS) that makes memory bases for DeepSeek Harness, Hermes, OpenClaw and other agents
I really want to brag about this new project that has just been open sourced. Probably one of the most worth installing local AI infrastructure this year, It solves a pain point problem that I have endured using AI Agent for more than half a year: Changing an AI tool requires re-explaining the project background, technology selection, and pitfalls.
You have reached the end of the archive
All of deepseek