ollama/ollama

↓ decelerating · ollama/ollama

Ollama decelerated to -7.7 with velocity dropping to 12.7 from 20.4 stars per day, post v0.21.0's Hermes Agent addition on April 16 and ROCm updates in v0.20.8, which enhanced AMD GPU compatibility but led to a saturation point. This trails llama.cpp's 31.0 but aligns with vLLM's -7.3, indicating a runtime-wide trend. For category evaluators, this means watching for MLX integrations to reverse the trend, as decelerations below -5 often correlate with 14-day velocity floors around 10 if unaddressed.

Receipts — documents this drew from

• Ollama (ollama/ollama)
• llama.cpp (ggml-org/llama.cpp)
• vLLM (vllm-project/vllm)

From the briefing: 2026-04-27 · Inference Runtimes Decelerate Amid Platform Acceleration