Oxmiq Open-Sources NVFP4 Quantized Hy4 Model, Runs on RTX 6000 Pro
RajaXg · x · 2026-09-30
- Oxmiq Labs released an NVFP4 (W4A16) quantized open-source version of Hy4-preview on Hugging Face.
- Built for Tencent's Hy4 model, the quantized variant targets RTX 6000 Pro server deployments and reportedly works well.
- The poster says the team benefits from open source and plans to contribute far more going forward.
- The model uses custom special tokens and chat template, supporting high/nothink reasoning modes and tool calls.
More from Infra
- PyTorchCon NA puts vLLM center stage: KV cache, disaggregated serving, expert parallelism — PyTorch · 2026-09-30
- Inference Engineering Without a GPU: Engineer Details a CPU-First Path into the Field — abhijithneil · 2026-09-30
- Anthropic carries $252B in non-cancelable compute obligations plus $161.2B in Broadcom TPU leases — ns123abc · 2026-09-30
- Datacenter backlash accelerating: indexed YouTube videos up 4.8x in 30 days — zealcaiden · 2026-09-30
- Blackstone, KKR, Apollo put $5.3B into gas plants powering AI data centers — pstAsiatech · 2026-09-30
- 'Not from a neocloud': engineers mock GPU elasticity gaps of small cloud providers — charles_irl · 2026-09-30