GLM-5.3-Flash vs Hy3

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.

Hy3

Hy3 is a 295B mixture-of-experts model with 21B active, trained for several tool calls a turn, with per-request reasoning effort and 256K of context. It needs an 8-GPU server for FP8 or about 200 GB for a 4-bit build. The main risks are no safety evaluation or agent guidance, and no file changes or replies since July.

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

Hy3

1. Start it with vLLM on 8 GPUs, vllm serve tencent/Hy3-FP8 --tensor-parallel-size 8 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3
2. In SGLang use --tool-call-parser hunyuan --reasoning-parser hunyuan, as the card shows
3. Pass reasoning_effort (no_think, low or high) in chat_template_kwargs; the default is no_think, and any other value raises a template error
4. With the vLLM reasoning parser the chain of thought comes back in message.reasoning and the answer in message.content
5. Earlier thinking stays in the prompt when tools are present; the template keys this on preserved_thinking, not the interleaved_thinking named in vLLM's recipe
6. Use temperature 0.9 and top_p 1.0, the card's recommended sampling
7. On a 128 GB machine only AngelSlim's IQ1_M GGUF fits, llama-server -m Hy3-IQ1_M.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536, with llama.cpp b11768 or later; at 89.4 GB × 1.2 it sits right at the 107.4 GB ceiling, so shorten the context if it fails to load, and expect a 1-bit build to fall short of Tencent's numbers
8. Without the hardware, call it through Featherless AI or DeepInfra via Hugging Face's inference providers