# NVIDIA TensorRT-LLM 1.4.0rc1 adds DeepSeek-V4.1-Flash, defaults to KV cache V2

> **Key takeaway:** NVIDIA's TensorRT-LLM inference engine reaches release candidate 1.4.0 with breaking API changes to KV cache management, new model support for DeepSeek-V4.1-Flash and GLM-5.3-Flash, and extensive Blackwell/Rubin architecture optimizations.

**Published:** 2026-10-22T01:38:13+00:00 
**Source:** NVIDIA TensorRT-LLM Releases 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/nvidia-tensorrt-llm-140rc1-adds-deepseek-v41-flash-defaults-.html 

## Summary
NVIDIA has published TensorRT-LLM v1.4.0rc1, a pre-release of its optimized large language model inference engine. The update makes KV cache manager V2 and the Python KV cache transceiver the default, a breaking change for existing deployments. New model support includes DeepSeek-V4.1-Flash, NVFP4-quantized GLM-5.3-Flash checkpoints, and experimental Qwen3.8-Flash-Next NVFP4 targeting. The release also introduces the Mooncake connector integration for disaggregated serving, adds runtime sleep/wakeup capabilities through the LLM API and trtllm-serve, and delivers numerous kernel-level optimizations for Blackwell and the upcoming Rubin architecture—including CuTe DSL MoE improvements, packed query heads for MLA, and locality-domain execution support.

## What's new in TensorRT-LLM 1.4.0rc1
- **Breaking API changes:** KV cache manager V2 and the Python KV cache transceiver are now the default, and unused `LlmRequest` fields have been removed, requiring code updates for existing integrations.
- **Model support expansion:** Adds DeepSeek-V4.1-Flash, NVFP4 GLM-5.3-Flash checkpoints, experimental Qwen3.8-Flash-Next NVFP4 as a Modeling V2 target, and LiLiCorr draft checkpoints in DFlash; removes automatic W4A16 fallback for Nemotron-H NVFP4 on Hopper.
- **KV cache and serving architecture:** Integrates the Mooncake connector with KV cache manager V2 for disaggregated serving, adds KV cache lending API for transfer backends, exposes native KV event streaming, and makes V2 pool rebalancing parameters configurable.
- **Blackwell and Rubin optimizations:** Adds Rubin locality-domain execution, SM107 support, packed query heads and compact inputs in Blackwell CuTe DSL MLA, PrimTS dense BF16/FP8 context attention paths, and SwiGLU clamping with updated NVFP4 FC1 weight layout in Rubin CuTe DSL MoE.
- **Multimodal and specific model optimizations:** Optimizes Kimi K3 on Rubin, speeds up MiniMax-H3 video-VAE decoding with parallel tiled decoding and Ulysses parallelism, adds token-sharded tensor parallelism for Wan in VisualGen, and unifies multimodal encoder caching to reduce output copies.

## Why it matters
TensorRT-LLM remains a widely deployed inference backend for production LLM serving on NVIDIA hardware. The breaking switch to KV cache manager V2 will force operators to validate and update their deployments before upgrading, while the Mooncake connector integration signals NVIDIA's deepening investment in disaggregated prefill-decode architectures that reduce GPU memory pressure. The extensive Blackwell and Rubin kernel work indicates this codebase is already being positioned for NVIDIA's next-generation accelerators, making this release relevant for both current Hopper/Blackwell operators and hardware planners evaluating future infrastructure.