Analyst: DeepSeek has post-trained V4 on Ascend since summer; real question is scaling to 5T+ params
AccBalanced · x · 2026-11-04
Chip analyst tphuang says the DeepSeek V5 + Ascend news is no surprise: he heard since summer that DS was post-training V4 on Ascend, and people who used the 950 consider it better than Hopper. He argues the discussion should shift from which chips DS uses to whether it can scale to 5T+ parameters, noting each base model brings significant challenges and V4 took a long time.
More from Infra
- Sipeed's 10cm MaixStation Dock claims DGX Spark-class local AI inference, runs off a power bank — vincent_koc · 2026-11-04
- vLLM, Red Hat and NVIDIA to host open-source inference meetup in Madrid on Nov 12 — vllm_project · 2026-11-04
- vLLM v0.33.0 ships with 927 commits, 1.5–2x faster decode on GB300 — vllm_project · 2026-11-04
- Fireworks launches Ultra-fast inference up to 1,500 tok/s and half-price background mode at Forge — nicolechirps · 2026-11-04
- Fireworks event slides show sliding-window bit-sharing quantization trick — sophiamyang · 2026-11-04
- ThunderAgent merged into LLM-D: 2x agent serving throughput, TTFT from ~50s to 2-4s — simran_s_arora · 2026-11-04