Ollama’s performance has fallen behind llama.cpp by a wide margin. Community benchmarks show Ollama delivering 30–70% fewer tokens per second on the same hardware and models. For example, on an RTX 5090 running Qwen3 Coder 32B, llama.cpp hit 52 tokens/sec, while Ollama managed just 30 tokens/sec. Georgi Gerganov, llama.cpp’s creator, pointed out that Ollama’s custom ggml kernels introduce extra branching and inefficient attention code. Users who switched back to llama.cpp report consistent throughput gains without any drop in output quality. Beyond speed, Ollama’s fork of ggml led to proprietary model storage and vendor lock-in. Between 2024 and early 2026, downloaded models lived in hashed filenames tied to Ollama’s registry format. You couldn’t reuse those weights with other tools. During that time, their inference engine was also slower. Although Ollama reverted to upstream llama.cpp in v0.30.0-rc15—citing new model demands, OpenAI’s GPT-OSS release, and community pressure—many feel the damage is done and trust is eroded. Ollama’s pivot toward a cloud platform further undermines its local-first promise. Ollama Cloud Pro users report failure rates up to 95%, long timeouts, broken tool calling and harsh rate limits that throttle paid users after days of use. Support tickets go unanswered for weeks. Meanwhile, privacy questions loom: shifting prompts to a cloud service contradicts the original sell of fully local inference. Finally, trust issues keep cropping up. Ollama obscured model names—labeling distilled variants of DeepSeek-R1 as the full 671 B-parameter model—leading to user confusion. Their desktop app launched with unclear licensing and no source code, despite claims of open-source status. As llama.cpp and other open-source projects continue rapid innovation, Ollama’s mix of slower performance, proprietary formats and shaky transparency makes sticking with it in 2026 a poor choice.