Tencent's Hy-Nomas world model needs at least 24GB VRAM, dev complains
cocktailpeanut · x · 2026-10-24
Pinokio's cocktailpeanut breaks down Tencent's newly released Hy-Nomas world model requirements: minimum is a single RTX 4090 (24GB) using the low-memory config — INT8 DiT weights and KV cache, FP8 PixelUMM decoder, sliding-window history, and CPU offloading of text encoder/DiT/VAE, which needs lots of system RAM and slows generation. Full-precision inference needs the FP32 distilled DiT with sequence parallelism (multi-GPU); LoRA training reference config uses 128 GPUs. His complaint: devs should optimize for consumer machines before release, not wait for community gguf ports.
More from Infra
- Baseten unveils Project Beacon to build safety tooling into the inference layer for open models — baseten · 2026-10-24
- Miles v0.1.3 cuts first RL weight update from 99.5s to 3.9s, shares GPUs across train/rollout — hsu_byron · 2026-10-24
- 98%+ Tensor Core utilization shown in LLM training run, drawing admiration — gharik · 2026-10-24
- We spent millions on AI apps with zero revenue — but our old GPUs are now worth more than we paid — stephen280up · 2026-10-24
- Polymarket puts 24% odds on AI data centers launching to space by end of next year — Polymarket · 2026-10-24
- Polymarket jokes wolves are entering Finnish villages after Google cleared forest for AI data centers — Polymarket · 2026-10-24