compare
KTransformers vs llama.cpp
The same facts for both, read from GitHub every night, and the relation a person reviewed.
| Fact | KTransformers | llama.cpp |
|---|---|---|
| Language | Python | C++ |
| Licence | Apache-2.0 | MIT |
| Stars | 20k | 132k |
| Latest | v0.7.2 | v0.7.0 |
| Last push | 2026-10-16 | 2026-10-19 |
| Release cadence | about 26 days between releases | too few releases |
| Active contributors | 12 commit authors on the default branch in the last 90 days | 175+ commit authors on the default branch in the last 90 days |
| Flags | none | none |
How they relate
KTransformers replaces llama.cpp. Alternatives to llama.cpp →
PartialRuns large mixture-of-experts models split across CPU and GPU, on top of its own optimized kernels.
KTransformers
- v0.7.22026-10-15KTransformers v0.7.2
- v0.7.12026-09-15KTransformers v0.7.1
- v0.7.0.post42026-09-13KTransformers v0.7.0.post4
- v0.7.02026-08-17KTransformers v0.7.0
- v0.6.42026-07-23KTransformers v0.6.4
llama.cpp
- b117232026-10-19hexagon: FLASH_ATTN_EXT improvements for longer context and SWA models (#30731)pre-release
- b117222026-10-19hexagon: allow multistream FA (#30218)pre-release
- b117212026-10-18cuda : fuse qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain (#30602)pre-release
- b117202026-10-18ggml: support non-contiguous SUM (CPU, CUDA, Vulkan, WebGPU) (#28851)pre-release
- b117192026-10-18ggml-cuda: fattn do not check oob if k_VKQ_max = (k_VKQ_max - nbatch_fa) (#30572)pre-release
KTransformers
✓ signed The latest release, v0.7.2, carries a signature GitHub verified.
Loading the security report
llama.cpp
unsigned The latest release, v0.7.0, carries no signature GitHub could verify.
Loading the security report