The new release adds support for several models, makes the KV cache manager V2 the default, and removes unused request fields.
NoteBy OossaPublished by Oossa: 1 min read
NVIDIA posted TensorRT‑LLM version 1.4.0rc1 on October 22 2026. The update makes the KV cache manager V2 the default API, a breaking change for developers, and drops the now‑unused LlmRequest fields. It also adds model support for DeepSeek‑V4.1‑Flash, NVFP4 GLM‑5.3‑Flash, and Qwen3.8‑Flash‑Next, among others.
Developers using TensorRT‑LLM must update their code to the new KV cache API, or their applications will fail to start.
Models
Flint offers 40 billion parameters, 2.5 billion active, and runs at 125 tokens/sec on a MacBook Pro.
Models
The new Flint model helps Ollama understand and modify code across large projects, with a 262K‑token window and early Mac support.
Models
Google's Gemini AI can now answer spoken navigation questions in Google Maps on CarPlay, starting in the United States.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
