Tencent Hunyuan Hy3

Tencent's entry in the open MoE race, Hy3 packs 295B parameters into 192 experts and routes each token through 8 of them, so only about 21B activate per step, while a 3.8B multi-token-prediction layer accelerates decoding through speculative execution. Context stretches to 256K tokens, a reasoning_effort switch moves between deep chain-of-thought and a no_think mode for direct answers, and tool calling is built in. The summer 2026 open-sourcing, following an April preview, includes BF16 and FP8 weights on Hugging Face and ModelScope under Apache-2.0, serving recipes for vLLM and SGLang with speculative decoding enabled, TensorRT-LLM support, a fine-tuning pipeline, GRPO reinforcement learning via the verl framework, and the AngelSlim quantization toolkit. Tencent recommends eight large-memory GPUs such as H20-3e class hardware for full-precision serving, placing self-hosting firmly in cluster territory, though hosted endpoints picked it up quickly and it climbed OpenRouter usage rankings soon after launch.