Hy3 vs Qwen3.8-27B

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.

Qwen3.8-27B

A dense cousin for coding and agent work with image and video input that runs almost anywhere. Ollama's 4-bit build needs about 22 GB, and five providers host it for agents without the hardware. The main risk is that nobody at Alibaba answers, there are no safety results or agent guidance, and the tool format is only in the template.

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

Qwen3.8-27B

1. Quickest start is ollama run qwen3.8, about 18 GB of files
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3.8-27B --max-model-len 131072 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, since FP8 on gfx1151 isn't documented
3. Use the qwen3_coder tool parser, SGLang's cookbook says the Hermes parser leaves its calls unparsed
4. Thinking is on by default, pass chat_template_kwargs with enable_thinking false for short tasks, and read the reasoning from reasoning_content
5. reasoning_effort takes xhigh (default), medium or low, and any other value makes the template raise an error
6. Sampling per the card is temperature 1.0 and top_p 0.95 when thinking, temperature 0.7, top_p 0.8 and presence_penalty 1.5 without
7. Preserved thinking keeps earlier reasoning in the context unless preserve_thinking is false, which grows long agent sessions
8. Turn on YaRN only beyond 262,144 tokens, the card warns static YaRN can cost quality on short texts
9. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented