Pangu vs DeepSeek V3 training cost gap: partly skill issue, partly method
teortaxesTex · x · 2026-11-21
A technical thread adds caveats to Pangu vs DeepSeek V3 training cost comparisons: Pangu is 505B params (18B activated) vs V3's 671B (37B), and may be inherently trickier to train. Pangu's step time also includes the costlier Muon optimizer plus mHC and SWA. The author still suspects the original figures may be a large overestimate.
More from Infra
- AI gateways are becoming the key control layer in production AI stacks: 8 capabilities — goyalshaliniuk · 2026-11-21
- Crowds line up to try local AI demos at NVIDIA RTX Spark launch — JFPuget · 2026-11-21
- Inkling-2-Mini quants are a trap: 17GB weights overflow 48GB VRAM at long context — EmilPi · 2026-11-21
- Cloudflare launches multimodal model clairic-flash, trending on Hugging Face — Cloudflare · 2026-11-21
- Hy AMA reveals Hy4 Flash runs on 96GB VRAM + 128GB RAM, plus QAT quantization details — allegedlyalienated · 2026-11-21
- Inkling-2-Mini runs on laptops with as little as 9GB RAM, per Unsloth local guide — NielsRogge · 2026-11-21