Tencent
Tencent's 295B mixture of experts — frontier-class reasoning, small-model economics.
262.144K tokens
Konteks
128,000 tokens
Output maks
Cepat
Kecepatan
Hard analysis at a cost that suits everyday use.
Read, explain and rewrite code across a large context.
Strong in English and Chinese in the same thread.
Efficient inference makes long sessions practical.
Tuned to this model — click any line to copy.
HY3 is Tencent's mixture-of-experts model, released in July 2026 and built for reasoning, agentic workflows and the kind of production traffic where cost per request matters as much as quality. On NinjaChat it is included in every plan, next to models from OpenAI, Anthropic, Google and Z.ai.
The architecture is the story. HY3 has 295 billion parameters in total but activates only about 21 billion for any given token, routing each one through the best 8 of 192 experts. That is why a model of its size can be served at a price closer to a small one, and it is the same design idea that made DeepSeek's V4 line competitive on cost.
It also exposes a configurable reasoning effort, so the model can think briefly on an easy turn and at length on a hard one rather than spending the same budget on everything. It handles about 262K tokens of context and is strong in both English and Chinese.
One honest limitation on NinjaChat: HY3 runs here without tool calling. Different hosts return tool arguments in inconsistent shapes for this model, so rather than ship something that breaks mid-agent-loop, it is served for reading, reasoning and writing. When a task needs function calls, GLM 5.3 and Kimi K3 sit in the same picker.
Dari nol ke hasil pertama Anda dalam waktu kurang dari satu menit.
01
Create a NinjaChat account and choose a plan
02
Open chat and pick HY3 from the model list
03
Give it the full problem; it has room for 262K tokens
04
Ask it to think longer on the turns that deserve it
05
Switch to GLM 5.3 or Kimi K3 when the task needs tool calls
Perbandingan jujur — di mana {model} unggul, dan di mana tidak.
Configurable reasoning effort and strong bilingual work
DeepSeek V4 Flash has a larger context and tool calling
Cheaper per message for everyday reasoning
GLM 5.3 is far stronger on hard engineering and supports tools
Better suited to focused reasoning tasks
MiniMax M3 reads images and video and writes much longer answers
Setiap paket NinjaChat sudah termasuk 50+ model, studio gambar lengkap, dan pembuatan video.



Satu langganan mencakup semua model di NinjaChat, termasuk HY3.