Ollama vs Pictory

The last ten releases are almost entirely runtime and backend work: NVFP4 kernel fusion for faster prefill, a repeat_penalty default change to match other engines, Laguna model support built on MLX and then handed back to upstream llama.cpp, CUDA compute-capability coverage for B200-class cards, and iGPU projector offload. Nearly every entry arrives as a release candidate; finals are rare enough that the v0.32.10 tag is the exception. User-facing surface area has barely moved.

Ollama is settling into a maintenance posture on the engine and pushing model-specific work upstream rather than carrying local forks — the Laguna implementation was added in one release and removed in favor of llama.cpp two days later. The remaining local investment is in Apple MLX quantization and hardware coverage, where being first to run a checkpoint on consumer silicon is the differentiator. Performance claims are now benchmarked and A/B verified in the notes, which is a change in rigor if not direction.

Expect the next releases to keep chasing new model families on MLX and to keep folding them upstream once llama.cpp catches up. The repeat_penalty default change is the kind of behavior shift that usually generates a follow-up fix once older models start repeating themselves in the wild.