vLLM v0.28.0rc1 Released: Security Hardening and Inference Optimization Deep Dive

vLLM releases v0.28.0rc1, focusing on remote code trust security hardening in the inference engine.
vLLM has released v0.28.0rc1, the final validation stage before the official stable release. The highlight of this update is a security fix: a `resolve_trust_remote_code` guard is introduced for `_load_ov2_processor`, ensuring third-party model remote code only executes with explicit user authorization — reducing supply chain attack risks. The article also revisits vLLM's core performance advantage via its PagedAttention memory management, which minimizes GPU memory fragmentation and boosts concurrent throughput. Developers are advised to test the RC in non-production environments and wait for the stable release before upgrading production deployments.
vLLM: A Continuously Evolving Open-Source Inference Engine
vLLM stands as one of the most popular inference and serving frameworks for large language models today. Known for its high throughput and low latency, it has earned over 90.9k stars on GitHub and has been forked by more than 21,000 projects. The vLLM team has recently released v0.28.0rc1 (Release Candidate 1), signaling that the latest round of feature development has entered its final validation phase.
For developers who closely follow LLM deployment and inference optimization, every vLLM release deserves attention. As an inference engine renowned for its PagedAttention mechanism, vLLM continuously refines performance, security, and model compatibility. While this rc1 release is still in the candidate stage, the commit history reveals the key directions driving this update.

Security Hardening: Strengthening the Remote Code Trust Mechanism
A clear theme emerges from the commit messages in this release — Bugfix and Security. Specifically, this version fixes the loading logic in _load_ov2_processor by introducing a resolve_trust_remote_code guard.
Why This Security Fix Matters
In the large model ecosystem, many models — especially third-party models from Hugging Face Hub — execute custom Python code from remote repositories when loading processors or tokenizers. This is the so-called trust_remote_code mechanism. While it enables flexibility, it also introduces a potential security risk: without proper constraints, remote code execution could allow a malicious model to run arbitrary code on your server.
By guarding _load_ov2_processor with resolve_trust_remote_code, vLLM effectively adds a security gate at the processor loading stage, ensuring that remote code is only executed when the user has explicitly granted permission. For service providers deploying vLLM in production — especially those that dynamically load a variety of models — this type of security hardening significantly reduces the risk of supply chain attacks.
Understanding Release Candidates: What RC Means for You
vLLM follows the industry-standard semantic versioning strategy. The v0.28.0rc1 release is a Release Candidate, meaning core functionality is essentially frozen and the team is conducting final stability testing and bug fixes.
How Developers Should Approach RC Versions
For users eager to experience new features early, RC versions offer a preview channel. However, it's worth noting that release candidates have not yet undergone full production-environment validation and may still contain undiscovered issues. The recommended approach is to differentiate usage as follows:
- Experimental and test environments: Actively try rc1 and help the community identify potential issues.
- Production environments: Wait for the official stable
v0.28.0release before upgrading. - Contribute feedback: If you encounter anomalies during testing, submit an issue to the vLLM GitHub repository — this kind of community participation is vital to the healthy development of open-source projects.
Why vLLM Continues to Deserve Your Attention
PagedAttention: The Core Technology Behind High-Performance Inference
What sets vLLM apart from other inference frameworks is its innovative PagedAttention memory management mechanism. Inspired by the concept of virtual memory paging in operating systems, this technique stores KV caches in non-contiguous memory blocks, dramatically reducing memory fragmentation and waste — and significantly boosting concurrent throughput as a result.
In practice, this means the same GPU hardware can serve far more concurrent requests, directly lowering the per-unit cost of large model inference. This is the primary reason so many enterprises and research institutions choose vLLM as their inference backend.
A Thriving Open-Source Ecosystem and Community
Over 90.9k stars and 21.7k forks speak directly to the vitality of the vLLM community. This scale not only means a wealth of use cases and documentation resources, but also ensures the framework can rapidly adopt the latest model architectures and optimization techniques. From mainstream open-source LLMs to the newest multimodal models, vLLM's compatibility continues to expand.
Summary and Upgrade Recommendations
Although vLLM v0.28.0rc1 is a release candidate, the security hardening it carries demonstrates the project team's sustained commitment to production readiness. As large models increasingly become core infrastructure, the security and stability of inference frameworks are becoming just as important as raw performance.
For developers and operations teams, closely tracking vLLM's version evolution not only helps capture performance optimization gains but also enables timely responses to security-level risks. We look forward to the official v0.28.0 stable release, which will allow for a more comprehensive evaluation of the overall improvements this update brings.
Tip: To try out or track this release, visit the official vLLM GitHub repository (vllm-project/vllm) for the complete release notes and changelog.
Related articles

DeepSeek V4 Pro Burning Through Credits Too Fast? The Hidden Logic Behind AI Model Pricing
Why does DeepSeek V4 Pro drain credits so fast while Flash barely moves? A deep dive into AI token billing, Pro vs. Flash pricing differences, and cost optimization tips.

RealPDE Competition Breakdown: The Frontier Challenge of AI-Powered Real-World Fluid Dynamics PDE Solving
A deep dive into the NeurIPS 2026 RealPDE Competition, covering the Sim2Real and LTTTA tracks, and how neural operators tackle real-world PIV and CFD fluid PDE challenges.

Building a Production-Grade 3DGS Training Library from Scratch: A Deep Dive into Full-GPU Residency and the Vulkan Stack
A veteran graphics engineer builds a production-grade 3DGS training library from scratch using C++23, CUDA, and Vulkan, achieving 60fps with 5M splats. Deep dive into its architecture and design.