A former Java colleague was forced by his company to pivot to Agent work (applications where AI autonomously completes tasks), and his very first question was "which large model is the strongest" — precisely the most common misjudgment in current Chinese AI engineering implementation. A Juejin engineering note cuts through this cleanly: first look at what the model does inside the system, then talk about parameters and scores.

What This Is

We note that the core of this note is pulling "large models" out of vague worship and classifying them along two dimensions.First, "modality" — what forms of information a model can receive and output — text, image, audio, and video each count as one. Multimodal does not mean "can do everything"; you have to check what's supported on input and output respectively.Second, "function" — the model's role in the Agent system, sorted into four categories:
  • Generation model: understands, reasons, and produces answers, playing the role of "brain."
  • Embedding model: converts documents into numerical vectors (so machines can compare semantic similarity), used for fast candidate recall from massive corpora.
  • Rerank model: fine-ranks initial retrieval results, filtering from "possibly relevant" to "definitely relevant."
  • Classification model: judges intent and risk levels, handles content routing.
The note contains a judgment we consider important: tool calling is the model generating parameters; the actual execution is done by application code — permissions, timeouts, and retries are system-layer concerns. This punctures the illusion that "a model that can call tools is an Agent."

Industry View

The mainstream narrative defaults to "the stronger the model, the more deployable the Agent," and everyone is chasing flagship benchmarks.This note represents the engineering camp's counter-position: splitting generation, retrieval, and ranking across different models yields lower overall cost, faster response, and more stable results. This is the same line as how RAG architecture (having AI look up information before answering) has gradually replaced the "stuff everything into context" approach in enterprises over the past two years.But the risks cannot go unmentioned. Multi-model deployment means four sets of operations, four sets of monitoring, and rising version-alignment costs. Embedding vector dimensions (BGE-M3 is 1024 dimensions, BGE Small is 384) are not the higher the better; what fits in storage and runs fast enough is the engineering reality. Using multiple vendors' models brings switching costs and data compliance risks that rise in tandem — for small and medium companies, this may not actually reduce burden.

Impact on Regular People

  • For enterprise IT: We recommend that AI procurement should not only盯 benchmark scores; first, clearly break down the tasks. Spending the same budget on "one super-large model doing everything" is often inferior to spending on "four cheaper models each handling a track."
  • For individual careers: Future AI invocation will be routed by scenario — the system that writes reports and the system that queries company policies may not be the same one. People who understand this division of labor are more valuable.
  • For the consumer market: An "all-purpose AI" won't arrive in the short term. The visible trend is each product becoming more focused: customer service AI more accurate, search AI faster, writing AI longer — because behind the scenes, different models each do their own work.