도구 호출과 오픈소스 모델의 M×N 문제

Tool calling with closed-source models is seamless. You pass a list of functions to the API, the model calls them, you get structured JSON back. The wire format is invisible to you. Then you move to open models and discover that tool calling depends on a wire format the engine has to understand. If the engine doesn’t support that model’s format yet, the output comes back garbled: reasoning tokens in arguments, malformed JSON, missing tool calls. Then you either wait, or write the parser yourself. What “supporting a model” actually means Every model family encodes tool calls differently. Here’s the same semantic operation, calling a function search(query="GPU") , in three wire formats: gpt-oss (Harmony): <|channel|>commentary to=functions.search <|constrain|>json<|message|> {"query": "GPU"} <|call|> DeepSeek: <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>search '''json {"query": "GPU"} ''' <｜tool▁call▁end｜><｜tool▁calls▁end｜> GLM5 : <tool_call>search <arg_key>query</arg_key><arg_value>GPU</arg_value> </tool_call> Same operation, incompatible wire formats: different token vocabularies, boundary markers, and argument serialization schemes. To return a nice array of JSON objects with the generated tool calls, you need to parse the model output back into a clean API response. In practice, each of the M applications (vLLM, SGLang, TensorRT-LLM, transformers, etc.) ends up writing custom parsers for each model it wants to support. And that is only half of the implementation burden. The pace of the problem Gemma 4 is a good illustration of the difficulty involved. Its <|channel> reasoning tokens get stripped by the decoder before the parser sees them ( vLLM #38855 ). Reasoning content can leak into tool-call arguments ( vLLM PR #39027 ). The model’s non-standard format was different enough that llama.cpp had to abandon its generic autoparser and build a dedicated implementation ( llama.cpp PR #21418 ). These are training-time format choices surfacing as parser bugs. Generic parsers are swimming against the current The natural response is to build a parser generic enough to handle all formats. Every engine has tried. A reasonable heuristic, say “find special tokens, extract JSON between them,” covers some formats well enough. But then Harmony routes through <|channel|> with a to= attribute, and GLM5 serializes arguments as <arg_key> / <arg_value> pairs instead of JSON at all. This is the fundamental problem: wire formats are training-time decisions, and nothing constrains them to a shared convention. The space of possible formats is open-ended, so a generic parser is trying to anticipate design choices that haven’t been made yet. That is why generic parsers help with the common cases but do not eliminate the per-model tail, where the hard bugs live: reasoning tokens leaking into arguments, decoders stripping special tokens before the parser sees them, end-of-generation signals colliding with content. The same model-specific format knowledge is also needed during generation, not just after the fact when parsing the result. That is where grammar engines enter the picture. The missing separation When a new model ships, work happens in two independent places. Grammar engines, like Outlines, XGrammar, and llama.cpp’s grammar support, need to know where to apply constraints during generation: which tokens mark the tool-call envelope, when to activate structured generation inside it, and when to leave the model unconstrained outside it. Output parsers inside vLLM, SGLang, TensorRT-LLM, transformers need to do the reverse: take the raw generated text and extract tool calls into a clean API response. They need the same format knowledge in reverse. These are different teams, different codebases, different release cycles. But the model-specific knowledge they need is the same: which tokens mark the boundaries, how arguments are serialized, where reasoning tokens can appear. Today each team reverse-engineers this independently from chat templates and (if they’re lucky) documentation. The result is N models × M implementations of the same format knowledge, developed in parallel with no shared contract. A new model ships, and grammar engine maintainers and inference engine maintainers both start the same reverse-engineering work from scratch. We have already seen the ecosystem converge on shared chat templates in Hugging Face, standardizing how prompts and turns are formatted. Tool calling needs the same kind of separation: not one wire format, but a shared declarative way to describe them. Until that exists, each new model will keep triggering the same reverse-engineering work across the stack. The separation that’s missing is extracting that shared format knowledge into configuration rather than code. A model’s wire format, its boundary tokens, its argument serialization, and its reasoning token behavior, should be a declarative spec that both grammar engines and parsers consume. The model changes, you update the spec. The grammar engine and the parser don’t move. I am Rémi Louf, CEO of dottxt . Follow @remilouf / @dottxtai for our work on structured generation and tool calling.