Open-weight local LLM

Hy3

Tencent Hy Team MoE model with 256K context, strong agent/productivity benchmarks and Apache 2.0 licensing. Practical only for very large local workstations via IQ1_M GGUF.

Large-memory workstation 128 GB RAM IQ1_M High-end local coding assistant
Parameters
295B (21B active)
Minimum RAM
128 GB
Model size
89 GB
Quantization
IQ1_M

Can Hy3 run locally?

Hy3 needs a serious workstation with large unified memory or high VRAM.

Search for hy3 in LM Studio or another GGUF-compatible runtime.

chatcodereasoningbeasttool-callinggeneral

Install path

01
Check RAM fitMinimum 128 GB RAM. Start with the IQ1_M quant.
02
Load the modelSearch hy3 in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • Official Tencent Apache 2.0 open-weight release
  • Large MoE model with 21B active parameters and 256K context
  • Strong public claims for coding, frontend, CI/CD, tool-calling and long-context reliability
  • Mature GGUF path through AngelSlim with llama.cpp, LM Studio, Ollama and OpenClaw examples
  • Useful high-end local alternative to much larger flagship open-weight models
  • Reasoning-effort modes support direct answers or deeper thinking-style responses

Limitations

  • The practical GGUF path is still huge: IQ1_M is about 89 GB and Q4_K_M is about 182 GB
  • Comfortable local use needs a 128GB+ memory workstation or large-GPU setup
  • IQ1_M is an extreme low-bit recipe, so quality should be validated before serious use
  • Full server-style deployment still expects multi-GPU vLLM or SGLang infrastructure

Best use cases

  • High-end local coding assistant
  • Agentic tool-use workflows
  • Long-context productivity analysis
  • Frontend and CI/CD engineering tasks
  • Private workstation reasoning
  • Chinese-English enterprise assistant experiments

Capability profile

speed
2
quality
9
coding
9
reasoning
9

Technical notes

Developer
Tencent Hy Team
License
Apache 2.0
Context window
262,144 tokens
Architecture
Sparse Mixture-of-Experts model with 295B total parameters, 21B active parameters, 192 experts, top-8 routing and an MTP layer.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Similar models to compare

Where to go next