GLM-5.3-Flash vs Qwen3-4B-Instruct-2507

GLM-5.3-Flash

An MIT-licensed multimodal agent and coding model with a 1M-token context and several tool calls per turn, served by vLLM, SGLang, llama.cpp and Transformers. Z.ai's smallest build is 328 GB in FP8, so it needs a multi-GPU server. The main risk is that nothing on safety or prompt injection is published and there is no way to report a flaw.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

GLM-5.3-Flash

1. Start it with vllm serve zai-org/GLM-5.3-Flash --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47 (vLLM 0.32.0 or later, about 386 GB of GPU memory per vLLM's recipe)
2. Use the glm47 tool-call and reasoning parsers in vLLM or SGLang; calls come back as tool_call blocks with arg_key and arg_value pairs
3. Pass chat_template_kwargs: {"reasoning_effort": "low"} for short tasks; anything other than low or high falls back to max
4. Don't rely on thinking: false from SGLang's cookbook, since the template has no such switch and always opens a think block
5. Set clear_thinking: true in chat use, since the default keeps earlier reasoning in the prompt
6. On a 128 GB machine only third-party low-bit builds come near fitting, none of them from Z.ai

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented