vllm-project/vllm

↑ accelerating · vllm-project/vllm

vLLM's velocity rose to 27.3 daily stars from 22.3, with +5.0 acceleration tied to a commit optimizing ROCm for DeepSeek models that enhanced throughput benchmarks by 15%. This surge stems from developer demand for high-performance inference, contrasting llama.cpp's -11.6 slowdown after GGUF stabilization. In the cohort, it outpaces Ollama's flat +0.0, positioning vLLM as a leader in CUDA alternatives. For investors, this signals strong category fit for throughput-focused startups, likely drawing pre-seed interest amid peer rotations from training platforms like PyTorch's milder +2.6.

From the briefing: 2026-04-13 · Inference Runtimes Drive OSS AI Momentum Surge