Strata 0.2.16 Runs Its Own Engine, Drops llama.cpp, Claims Identical Output

Published by

Strata 0.2.16 introduces a new inference engine developed from scratch, eliminating the reliance on ggml and llama.cpp's matrix kernels, and claims to produce identical results across nine different GPUs. This update not only addresses a silent correctness bug affecting Apple M1 through M4 chips but also speeds up processing for long prompts on older NVIDIA cards. The developer, Niko Veit, emphasizes empirical testing to validate the output consistency and has included detailed performance metrics, showcasing significant speed improvements and robust platform fixes. Additionally, the new version supports a complex model called Qwen3.8-Flash-Next, which efficiently utilizes system resources, although it has known issues with memory constraints on certain setups



Strata 0.2.16 Runs Its Own Engine, Drops llama.cpp, Claims Identical Output

Strata 0.2.16 runs its models on an inference engine it wrote from scratch, dropping ggml and llama.cpp's matrix kernels across NVIDIA, AMD, Intel Arc and Apple Silicon. The bold central claim is that the rewrite produced identical answers to the old engine across nine different GPUs, verified against a wide spread of hardware. Alongside that, the update fixes a silent correctness bug that was producing wrong results on Apple M1 through M4 chips, speeds up long prompts on older NVIDIA cards, and folds in a pile of hard-won platform fixes.

Strata 0.2.16 Runs Its Own Engine, Drops llama.cpp, Claims Identical Output @ Linux Compatible