21B active parameters (295B total MoE) — punches well above its active parameter count, competitive with dense 70B+ models while keeping inference costs low
256K context, strong tool calling, configurable reasoning depth — covers the range from fast chat to long-context agent tasks
I think it’s worth adding, since it seems to run faster than a lot of the open-source models in its size class. I also find it has better design taste than most other models, and in general is worth adding. (definitely more deserving than Ling-2.6-Flash)
Log in to comment and vote
Comments1
Jade Custard
Adding my support for Hy3. It's already available on OpenRouter at openrouter.ai/tencent/hy3.
Apache 2.0 license — no commercial restrictions
21B active parameters (295B total MoE) — punches well above its active parameter count, competitive with dense 70B+ models while keeping inference costs low
256K context, strong tool calling, configurable reasoning depth — covers the range from fast chat to long-context agent tasks
I think it’s worth adding, since it seems to run faster than a lot of the open-source models in its size class. I also find it has better design taste than most other models, and in general is worth adding. (definitely more deserving than Ling-2.6-Flash)