# vLLM’s RC3 Release Lists Reflection Beam-1.0 and Muse Spark Open

Published 2026-11-02 · AI.info Team · AI.info

> The v0.33.0rc3 release lists support for Reflection Beam-1.0, Muse Spark Open and other models, alongside changes to CUDA, ROCm and PyTorch requirements. GitHub titles the page v0.33.0 and marks it Latest, while the release tag remains v0.33.0rc3.

[vLLM](https://github.com/vllm-project/vllm/releases/tag/v0.33.0rc3)’s third release candidate for version 0.33.0 adds support for Reflection Beam-1.0, Muse Spark Open and several other newly listed models. The project published the candidate on November 2, alongside changes to its PyTorch and CUDA requirements. The release notes describe a broad software update: 927 commits from 389 contributors, including 122 new contributors.

The model additions stand out for their range, from large mixture-of-experts systems to a visual document retriever. The release page is titled v0.33.0 and marked Latest by [GitHub](https://ai.info/companies/github), but the tag identifies it as [v0.33.0rc3](https://github.com/vllm-project/vllm/releases/tag/v0.33.0rc3). The page’s title and badge do not change the candidate suffix in the tag. The notes give model and backend details, while leaving users to check their own hardware and deployment needs before upgrading.

## Beam-1.0 brings a 501-billion-parameter model

Reflection Beam-1.0 is the largest model specified in the new-support list: a 501-billion-parameter mixture-of-experts model with 23 billion active parameters. vLLM lists day-0 support for the model, its DFlash2 speculative-decoding drafter and fused kernels for [NVIDIA](https://ai.info/companies/nvidia) Blackwell GPUs. The release also adds reasoning and tool parsers for Beam.

The candidate lists Yandex AliceAI-Foundation-80B-A3B, described as a hybrid model with Kimi Delta Attention, 512 experts and a 262,000-token context. Recursive Flint adds BF16 and NVFP4 GGUF checkpoint support; its listed configuration has 40 billion total parameters and 2.5 billion active. The notes also name Muse Spark Open and Step-5 Preview, among other model additions.

## Support extends beyond text generation

The list includes NVIDIA Alpamayo 2 Super, Nemotron-Labs Dory1.0-4B, Nemotron Parse 3.5, MiniCPM-OCR 1.0, Domyn Edge and ColUltraVBERT, a visual document retriever. vLLM says LFM2.5-Embedding-350M now uses its Lfm2BidirectionalModel implementation, while Kolibri 1 runs through the Transformers modeling backend. Decision 3.0 models and Cloudflare Clef-Flash are listed for the project’s `/v1/systemone` endpoint.

That mix makes the update relevant to teams serving more than chat models: it covers embedding, retrieval, document-image processing and structured decision tasks as well. The notes identify added engine support, but do not claim that every model runs on every device or configuration. Operators will still need to match model requirements to their GPUs and software stack.

## CUDA 12.9 users face an upgrade choice

vLLM 0.33.0 moves its CUDA wheels and images to CUDA 13.0 and drops the CUDA 12.9 packages. The release notes list ROCm 10.0 as the default ROCm option, while continuing to publish ROCm 7.2 packages. They also announce a CUDA wheel with a stable libtorch ABI and a new [AWS](https://ai.info/companies/amazon) EFA image for multi-node serving on eligible EC2 instances.

The project upgrades to PyTorch 2.14.0, torchvision 0.29.0 and Triton 3.8.0. For users evaluating RC3, those runtime shifts sit alongside the new model implementations: adoption depends not just on choosing a supported model, but also on checking CUDA or ROCm compatibility and any changes in the release notes. The candidate’s CUDA artifacts target version 13.0, not 12.9.

## Sources

1. Original source · Release notes [vLLM on GitHub](https://github.com/vllm-project/vllm/releases/tag/v0.33.0rc3) github.com

## Companies in this story

- [vLLM Project](https://ai.info/companies/vllm-project)

## Related stories

- [vLLM 0.33 Drops CUDA 12.9 Builds](https://ai.info/news/vllm-033-drops-cuda-129-builds)
- [vLLM details opt-in FlyDSL MonoKernels for faster decode on AMD GPUs](https://ai.info/news/vllm-flydsl-monokernels-agent-inference)
- [Three High-Severity vLLM Flaws Expose Serving Systems to Remote Attacks](https://ai.info/news/vllm-high-severity-flaws-serving-systems)
- [RadixArk team describes serving Meta’s Muse Spark Open on one TPU host](https://ai.info/news/pytorch-muse-spark-one-tpu-host)
- [Ai2 Releases AstaBrief, an 8B Model for Faster Scientific Reports](https://ai.info/news/ai2-releases-astabrief-8b-scientific-reports)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/vllm-0330-rc3-new-model-support
