This model was retired on 30 October 2026 and no longer accepts API requests. Recommended replacement: DeepSeek V4.1 Flash.
Hy3-FP8
Retired
Tencent's 295B-parameter Mixture-of-Experts reasoning model with just 21B active parameters — rivalling flagship open models that carry 2-5x more weights.
Quantization
FP8
MoE
Context Window
131K
Tokens
Intelligence
25.3
AA Index v4.3
Parameters
295B
21B active
Flagship quality, sparse compute
Hy3 is a 295B-parameter Mixture-of-Experts model with 21B active parameters, developed by the Tencent Hy Team. It outperforms similarly sized models and holds its own against flagship open-source releases carrying 2-5x more parameters.
Because only a small slice of the network fires per token, Hy3 is unusually cheap to serve — which makes it a strong default for reasoning-heavy async and batch workloads where you care about cost per million tokens as much as answer quality.
295B total · 21B active per token
Best for cost-sensitive reasoning at volume
Efficient Reasoning
Only 21B parameters active per token, so you get flagship-class reasoning quality at a fraction of the serving cost.
High-Volume Pipelines
Cheap enough to run across millions of documents, tickets, or records in async and batch jobs.
Tool & Agent Steps
Reliable instruction following makes Hy3 a strong workhorse for routing, extraction, and multi-step agent stages.
Knowledge Work
131K context handles long reports, transcripts, and codebase excerpts in a single call.