GLM-5.3 vs Qwen3.5-2B

GLM-5.3

A large coding and long-horizon agent model with a 1M-token context, several tool calls per turn and three reasoning budgets. Z.ai's smallest build is 756 GB in FP8, so it needs an 8-GPU server, Z.ai's API or a provider. The main risks are no published safety work, no way to report a flaw, and thinking that can't be turned off.

Qwen3.5-2B

A small vision-language cousin for prototypes, fine-tuning and light local steps such as tool calls, tagging or reading an image, running in about 2 GB at 4 bits in every major runtime under plain Apache 2.0. Keep thinking off, since the card warns it loops. The main risk is neglect, with no safety results or answers from Alibaba since March.

GLM-5.3

1. Start it with vllm serve zai-org/GLM-5.3 --kv-cache-dtype fp8 --tensor-parallel-size 8 --speculative-config.method mtp --speculative-config.num_speculative_tokens 5 --tool-call-parser glm47 --reasoning-parser glm47 --enable-auto-tool-choice --served-model-name glm-5.3 (vLLM 0.28.0 or later, 8xH200 or 8xH20 per vLLM's recipe)
2. Use the glm47 tool-call and reasoning parsers in vLLM or SGLang; SGLang's cookbook says glm45 leaves the calls as plain text
3. Pass chat_template_kwargs: {"reasoning_effort": "low"} for short tasks; anything other than low or high falls back to max
4. Pass clear_thinking: true in chat use, since the template keeps earlier reasoning by default
5. Don't send enable_thinking: false; the template ignores it and always opens a think block
6. Validate JSON output yourself, since structured output is documented only for Z.ai's hosted API
7. No Z.ai build fits a 128 GB machine; the full 1M context needs 8xB200 per vLLM's recipe

Qwen3.5-2B

1. Smallest start is ollama run qwen3.5:2b-q4_K_M (1.9 GB); plain qwen3.5:2b pulls the 2.7 GB Q8_0 build
2. Full precision vllm serve Qwen/Qwen3.5-2B --port 8000 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder
3. Thinking is off by default; if you turn it on with "chat_template_kwargs": {"enable_thinking": true}, stream and cut off runaway output, since the card warns of thinking loops, and add --reasoning-parser qwen3 in vLLM or SGLang to split the reasoning out
4. There's no generation_config.json, so pass the card's sampling yourself, for example temperature=1.0, top_p=1.0, top_k=20, presence_penalty=2.0 for text without thinking
5. Use it for narrow steps or as a fine-tuning base, not as the main agent, as the card itself suggests
6. Validate any JSON it returns, since structured output isn't documented