Models & Labs
vLLM v0.33.0: Blackwell & DeepSeek V4.1 Optimizations
vLLM Releaseshigh confidence
Why it matters
- →vLLM now supports CUDA 13.0 by default, essential for running models on new Blackwell hardware.
- →DeepSeek V4.1 and Reflection Beam-1.0 receive day-one optimization support for better performance.
- →AMD ROCm 10.0 becomes the default, improving parity with NVIDIA for open-source inference.
vLLM has released version 0.33.0, a significant update focusing on next-generation hardware support and performance optimizations for large language models. The release drops CUDA 12.9 in favor of CUDA 13.0 as the default, aligning with NVIDIA's Blackwell architecture, while also adding native support for DeepSeek V4.1 and Reflection Beam-1.0. Key technical improvements include block-diffusion drafters, quantized LM heads, and AMD ROCm 10.0 defaults, which collectively enhance decode throughput and interactivity on both NVIDIA and AMD GPUs. The update also removes legacy components like Medusa and the MLP speculator to streamline the codebase.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1llama.cpp b9952 Release Enhances DeepSeek V4
llama.cpp Releases · July 11, 2026 · Related
- 2vLLM v0.28.0 Release: Major Performance Enhancements
vLLM Releases · August 28, 2026 · Same story
- 3llama.cpp b10996 adds CUDA 13 and ROCm 10.0 builds
llama.cpp Releases · September 16, 2026 · Related
- 4vLLM v0.31.0: SM100 defaults and fast restarts
vLLM Releases · October 6, 2026 · Same story
More in Models & Labs
Models & Labsother
llama.cpp b12054 adds CUDA 13 and ROCm 10.0
This release quietly cements llama.cpp as the universal inference runtime by adding default support for CUDA 13 and ROCm 10.0. NVIDIA users can now leverage newer driver stacks without waiting for separate builds, while AMD GPU owners finally get parity with modern ROCm versions. The inclusion of Snapdragon and OpenVINO binaries further widens the hardware net, though Apple Silicon KleidiAI is notably disabled here. It’s less about new features and more about ensuring no major accelerator is left behind in the local LLM ecosystem.
llama.cpp Releases
Models & Labsmodels
llama.cpp adds Clover MoE model support
llama.cpp quietly expands its architecture roster by adding native support for the Clover Mixture of Experts model. This isn't just a wrapper; it implements causal hybrid attention and reuses existing recurrent-state and LoRA helpers, meaning local inference for this specific architecture is now viable on consumer hardware. The release includes GGUF conversion tools and loads across all major backends including CUDA 13, ROCm 10, and Apple Silicon. For researchers tracking non-transformer or hybrid architectures, this closes a gap in the most popular local inference runtime.
llama.cpp Releases
© TechCrunch AIModels & Labsother
Apple Watch Audio Intelligence Beta Begins
Apple is finally testing the AI features that define its wearable strategy. Live Rewind and Siri Recaps turn the Apple Watch into a passive conversation recorder, using on-device detection to buffer audio before sending it to Private Cloud Compute for summarization. The privacy architecture is aggressive: raw audio never leaves the secure enclave, and summaries are ephemeral by default. This marks the first time ambient AI recording has moved from prototype to consumer beta on a mainstream device.
TechCrunch AI