NVIDIA announced TensorRT‑LLM version 1.4.0rc0 on Oct 14, 2026. The update brings new model support like GLM‑5.3‑Flash and Qwen‑Image‑Layered, and introduces a sparse key‑value cache with tier‑aware locking. It also adds a Mooncake store pool and a V2 scheduler preemption feature. Several bugs were fixed, including hangs in multi‑GPU streaming on H100 GPUs and crashes with DeepSeek models on Blackwell GPUs.
Why it matters
Developers can run larger language models on NVIDIA GPUs with lower memory use and fewer crashes.
