Pithos v0.5.2

Pithos is a dependency-free C engine that runs Liquid AI's LFM2.5-24B-A2B and LFM3-1.7B locally,
with a terminal chat, the Pandora web chat, an OpenAI-compatible server and a terminal agent.
This release fixes --bench with long prompts; everything else works as in v0.5.1.

Downloads

File	For
pithos-0.5.2-windows-x64.zip	Windows 10 / 11, 64-bit (AVX2, AVX or scalar kernels chosen at run time)
pithos-0.5.2-linux-x64.tar.gz	Linux x86-64 with glibc 2.34 or newer (Ubuntu 22.04+, Debian 12+, RHEL / Rocky 9+, Fedora 35+...)
pithos-0.5.2-linux-arm64.tar.gz	Linux ARM64 (experimental): pithos for ARMv8.2 with dot product (Raspberry Pi 5, RK3566 / RK3588, Graviton 2+), pithos-armv8 for older cores (Raspberry Pi 3 / 4)
SHA256SUMS	checksums of the archives

The model is not included: download it as described in the README (LFM3-1.7B is 1.1 GB, the 24B
model 14.5 GB), put it in a models folder next to the program, then run for example
pithos --server --open.

The Windows binary is not signed, so SmartScreen may warn about an unknown publisher. The ARM64
binaries passed the kernel tests and a short generation on GitHub's ARM64 runners (Neoverse); they
have not been tried on single-board computers yet. Reports are welcome.

What's new since v0.5.1

• --bench with prompts longer than the context: the prompt pass was skipped and reported as
  0.00 s (thanks to @robertosantosx2 for spotting it in #2). The context now grows to fit the
  benchmark; with an explicit --ctx that is too small, --bench stops with an error instead.
  Chat, server and agent were not affected.

See the README for the full documentation and the measured results.

Pithos v0.5.1

Pithos is a dependency-free C engine that runs Liquid AI's LFM2.5-24B-A2B and LFM3-1.7B locally,
with a terminal chat, the Pandora web chat, an OpenAI-compatible server and a terminal agent.
This release makes the Vulkan backend faster on Intel GPUs; everything else works as in v0.5.0.

Downloads

File	For
pithos-0.5.1-windows-x64.zip	Windows 10 / 11, 64-bit (AVX2, AVX or scalar kernels chosen at run time)
pithos-0.5.1-linux-x64.tar.gz	Linux x86-64 with glibc 2.34 or newer (Ubuntu 22.04+, Debian 12+, RHEL / Rocky 9+, Fedora 35+...)
pithos-0.5.1-linux-arm64.tar.gz	Linux ARM64 (experimental): pithos for ARMv8.2 with dot product (Raspberry Pi 5, RK3566 / RK3588, Graviton 2+), pithos-armv8 for older cores (Raspberry Pi 3 / 4)
SHA256SUMS	checksums of the archives

The model is not included: download it as described in the README (LFM3-1.7B is 1.1 GB, the 24B
model 14.5 GB), put it in a models folder next to the program, then run for example
pithos --server --open.

The Windows binary is not signed, so SmartScreen may warn about an unknown publisher. The ARM64
binaries passed the kernel tests and a short generation on GitHub's ARM64 runners (Neoverse); they
have not been tried on single-board computers yet. Reports are welcome.

What's new since v0.5.0

• Vulkan on Intel GPUs: Intel devices get their own build of the four matmul kernels
  (other GPUs and Direct3D 12 run exactly the same code as before). LFM3-1.7B on an Intel UHD
  Graphics (i5-10500H): prompt 35 -> 53 tokens per second, generation 8.4 -> 10. Same results
  (perplexity unchanged); PITHOS_VK_INTEL=0 switches back to the common kernels.
• On that UHD the CPU is still faster (85 / 21), so --gpu auto keeps leaving integrated GPUs
  off; use --gpu vulkan --gpu-device N (see --list-gpus) to run on one.

See the README for the full documentation and the measured results.

Pithos v0.5.0

First release with prebuilt binaries. Pithos is a dependency-free C engine that runs Liquid AI's
LFM2.5-24B-A2B and LFM3-1.7B locally, with a terminal chat, the Pandora web chat, an
OpenAI-compatible server and a terminal agent.

Downloads

File	For
pithos-0.5.0-windows-x64.zip	Windows 10 / 11, 64-bit (AVX2, AVX or scalar kernels chosen at run time)
pithos-0.5.0-linux-x64.tar.gz	Linux x86-64 with glibc 2.34 or newer (Ubuntu 22.04+, Debian 12+, RHEL / Rocky 9+, Fedora 35+...)
pithos-0.5.0-linux-arm64.tar.gz	Linux ARM64 (experimental): pithos for ARMv8.2 with dot product (Raspberry Pi 5, RK3566 / RK3588, Graviton 2+), pithos-armv8 for older cores (Raspberry Pi 3 / 4)
SHA256SUMS	checksums of the archives

The model is not included: download it as described in the README (LFM3-1.7B is 1.1 GB, the 24B
model 14.5 GB), put it in a models folder next to the program, then run for example
pithos --server --open.

The Windows binary is not signed, so SmartScreen may warn about an unknown publisher. The ARM64
binaries passed the kernel tests and a short generation on GitHub's ARM64 runners (Neoverse); they
have not been tried on single-board computers yet. Reports are welcome.

What's new since v0.4

• Vulkan backend (Windows, Linux; experimental): the whole forward pass of dense models on the
  GPU. LFM3-1.7B on a GTX 1650: 328 tok/s of prompt, 82 of generation.
• Direct3D 12 backend (Windows, Xbox): the same for Direct3D 12 (330 / 69 on the same GPU).
• One option for the GPU: --gpu auto|off|vulkan|d3d12|opencl|opencl+vulkan|opencl+d3d12
  and --cpu. By default LFM3 runs on Vulkan on a discrete NVIDIA GPU and the 24B model on OpenCL;
  --list-gpus lists the devices of every backend, --gpu-device N picks one. The old options
  (--no-gpu, --d3d12, --vulkan) still work.
• LFM3-1.7B support: sliding-window and gated attention, the LFM3 chat template and tool calls.
• Reasoning levels: --reasoning high|medium|low|off or a token budget, also per request.
• AVX kernels for x86-64 CPUs without AVX2.
• Xbox One app (built from source, see the README): the server on a console in Developer Mode.
• Fixes: after a GPU failure the engine continues on the CPU from an empty context instead of
  returning blank answers; the server reports the Vulkan backend.

See the README for the full documentation and the measured results.