Hunyuan 3 (Hy3)

1 MLX alias · Tencent Hunyuan 3 · 295B MoE · preview · M3-Ultra-class (256 GB+) · hy_v3 tool + reasoning parser.

Hunyuan 3 (Hy3) is Tencent's 295B mixture-of-experts flagship, shipped in MLX form as a preview. It uses a dedicated hy_v3 tool-call parser and hy_v3 reasoning parser, and honors a reasoning_effort override on /v1/responses and /v1/chat/completions. This is an Ultra-class model — the weights want an M3 Ultra with 256 GB+ of unified memory, so it is not something a 32 or 64 GB Mac can hold.

Not R2-mirrored. Hy3 is not on the models.rapidmlx.com RAM picker or the R2 catalog — it pulls from Hugging Face on first use via its alias. Run rapid-mlx pull hy3-preview-4bit (or just serve) and the weights stream from mlx-community/Hy3-preview-4bit.

family

Hunyuan

aliases

1

status

preview

min RAM

256 GB (M3 Ultra-class)

install

rapid-mlx serve <alias>

OpenAI base URL

http://localhost:8000/v1

Hunyuan 3 295B · 1 alias · preview

295B MoE in 4-bit MLX. Dedicated hy_v3 tool-call parser and reasoning parser, with a reasoning_effort override. Ultra-class: plan for 256 GB+ of unified memory.

tool parser: hy_v3 · reasoning parser: hy_v3

Pull & serve

Notes & caveats

• Preview. Hy3 ships as a preview alias — expect the alias name and defaults to firm up as the family matures.
• Ultra-only. The 295B MoE weights need ~256 GB of unified memory; this is an M3 Ultra-class model, not a laptop model.
• Not R2-mirrored. Weights pull from Hugging Face (mlx-community/Hy3-preview-4bit), not the R2 mirror, so first run downloads from HF.
• Reasoning streams via the hy_v3 reasoning parser into reasoning_content; pass reasoning_effort to tune depth.
• Speculative decoding is not supported for this alias.

Where next