Apertus v1.5 8B vs Hy3

Apertus v1.5 8B

A fully open 8B cousin, training data included, for multilingual chat and tool calls with images and audio in. It needs about 22 GB in BF16 and Swiss AI's own vLLM or Transformers fork, since no released runtime supports it. The main risks are a gated download with a policy to pass on and tools that fail with thinking on.

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.

Apertus v1.5 8B

1. Accept the terms on the card and set HF_TOKEN before downloading
2. Use Swiss AI's Docker image ghcr.io/swiss-ai/vllm_apertus_1.5_release, since stock vLLM and Transformers don't load Apertus1p5ForConditionalGeneration yet
3. Serve with --chat-template-content-format string --enable-auto-tool-choice --tool-call-parser apertus for tools, and leave thinking off when you need tools
4. For thinking, set --reasoning-parser apertus and --default-chat-template-kwargs.enable_thinking true, and budget max tokens generously
5. Audio is resampled to 24 kHz and costs 40 tokens a second, and video must be passed as frames
6. Don't cast the loaded model with .half(), the image and audio tokenizers must stay in F32
7. If you redistribute the weights, the policy says you must pass it on and get the recipient's express assent

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