Repository navigation
hob-llama.cpp b11434: upstream llama.cpp (2026-10-05, c06f841) + Hob-forge's model ports, prebuilt.
Use this build to run Hob-forge GGUFs whose architectures aren't in upstream llama.cpp yet:
- Kolibri-1 (Aleph Alpha,
kolibri1): https://huggingface.co/Hob-forge/Kolibri-1-GGUF - BAR / FlexOlmo (AllenAI,
flex_olmo): https://huggingface.co/Hob-forge/BAR-2x7B-Tool-Use-GGUF and https://huggingface.co/Hob-forge/BAR-5x7B-GGUF
Everything else upstream llama.cpp supports works as normal.
Which file?
| You have | Download |
|---|---|
| Windows, NVIDIA GPU | win-cuda-12.4-x64 (older drivers) or win-cuda-13.4-x64 (newest drivers). Needs the NVIDIA CUDA runtime DLLs installed (CUDA Toolkit, or the cudart zip from an upstream llama.cpp release). |
| Windows, AMD GPU | win-rocm-10.0-x64 |
| Windows, any GPU (AMD, Intel, NVIDIA) | win-vulkan-x64 |
| Windows, CPU only | win-cpu-x64 (or win-cpu-arm64 for ARM laptops) |
| Linux | bin-ubuntu-x64 (CPU), bin-ubuntu-cuda-12.8-x64 (+ cudart-... if you don't have CUDA 12.8), bin-ubuntu-vulkan-x64, bin-ubuntu-rocm-10.0-x64, bin-ubuntu-arm64 |
| Mac (Intel) | bin-macos-x64 |
| Apple apps | xcframework |
Quick start (Kolibri-1, CPU):
llama-server -m Kolibri-1-Q4_K_M.gguf -c 32768 --jinja --temp 1.0 --top-p 0.97 --top-k 128
Then open http://127.0.0.1:8080. The architecture supports up to 262,144 tokens of context (-c 262144).
What's in it. The patches are on branch hob; see PORTING.md there.
- Kolibri-1, 4 commits: SIGMOID_LOGIT_ADD expert gating, converter, model graph, arch tests.
Verified against Aleph Alpha's reference: Q8_0 matches 67/67 top-1 tokens. - FlexOlmo, 1 commit. Matches the earlier April build byte-for-byte on a 32-token greedy test.
- 135/135
test-llama-archspass. - CI fork note: the s390x build was dropped because its runner only exists upstream.
Not done here: no Apple-silicon (Metal) macOS binary in this release; build from source on M-series Macs.
Licence: llama.cpp is MIT (see LICENSE-llama.cpp-MIT in each archive). Our patches are MIT too. This is an independent build, not made or endorsed by the llama.cpp project.
SHA256SUMS.txt lists checksums for every file.