Game SRE in Shanghai, ex-Tencent IEG. Running AI coding agents in production — writing what actually breaks. First-hand notes from China's game + AI stack.
Everyone's framing DeepSeek Harness as a Claude Code alternative.
Wrong level.
Claude Code and Codex are optimizing a coding agent. Harness modularizes the entire runtime — models, tools, skills, sessions, sandboxes, storage, even the agent loop itself. All plugins. All swappable.
The bet: the model isn't the moat. The harness is.
Still early. Still rough. But this is an OS for agents, not another coding tool.
China’s biggest tearjerker this year is Dear You — literally “Love Letter to Grandma.”
This week Justin Sun published a 10k-word “fiction” about a movie star who made him call her “mom.”
Chinese internet renamed it Love Letter to Mom. Same title. One character off.
RMB 140,000 for a Mac with 256GB of unified memory in China.
That’s literally car money.
And Apple still hasn’t priced the 512GB version.
512GB though? Local LLM freedom. Huge context windows. Code without OOM anxiety.
A car gets you places. This gets your tokens places.
Not even glazing GLM 5.3 here.
Fable 5 broke the job into 3 tasks. DS V4 Flash and Gemini 3.7 Flash both needed a second pass. GLM 5.3 one-shotted it — just slow as hell.
Grok 4.6 got lobotomized, so I’m back on the model market.
DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀
🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8.
Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model.
1/n