Blame for README.md

---
license: other
license_name: mixed-mit-apache-2.0
license_link: LICENSE
tags:
  - runtime
  - llama.cpp
  - llama-server
  - stable-diffusion.cpp
  - tts
  - text-to-image
  - image-generation
  - on-device
  - openbloom
---

# OpenBloom Runtime Releases

This repository hosts **OpenBloom's self-built runtime artifacts** — the engines and
bundles that the OpenBloom desktop application provisions at install time. It contains
**no model weights**: only the *runtime tooling* that serves open-weight models locally.
Every artifact is hash-pinned by the OpenBloom downloader and verified fail-closed on
every download.

## Contents

| Family | Path | What it is |
|---|---|---|
| **llama-server** | `llama-server/b12110/` | Our CUDA 12.8 build of [llama.cpp](https://github.com/ggml-org/llama.cpp) `b12110` (v0.9.0), merged with its cudart bundle as one flat artifact. The chat + ASR sidecar. |
| **sd-server** | `sd-server/5dcfa79/` | Our CPU build of [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp) at commit `5dcfa79` (`master-1104-5dcfa79`). The image-generation sidecar: SD/SDXL/Flux-class diffusion models served over a loopback HTTP API. |
| **KrisKross** | `kriskross/1.0.0 … 2.1.1/` | OpenBloom's build of the MOSS-TTS inference runtime: the OpenMOSS llama.cpp fork + the MOSS Audio Tokenizer (ONNX). CPU and CUDA 12 variants. The served GGUF weights live in their own model repo: [OpenBloom/MOSS-TTS-Realtime-GGUF](https://huggingface.co/OpenBloom/MOSS-TTS-Realtime-GGUF). |

## Verify before use

Every bundle ships a `MANIFEST.json` (the sha256 of each component) and every family's versions
carry a `SHA256SUMS` file. The consuming installer pins a sha256 and fails closed on any drift —
you can verify the same way:

```
# llama-server (CUDA 12.8, b12110)
llama-server/b12110/llama-server-b12110-linux-amd64-cuda12.tar.gz
sha256: 546dab15b51f86a43edd6452a8025e6fba425522102dbb6267cb1a376f7ad8dc

# sd-server (CPU, stable-diffusion.cpp 5dcfa79)
sd-server/5dcfa79/sd-server-5dcfa79-linux-amd64-cpu.tar.gz
sha256: a836547d079fb9c3236d3a6b8344fb93f431a478eb7fe8ba4daf54bfc00c8719

# KrisKross (latest, 2.1.1, CPU)
kriskross/2.1.1/kriskross-2.1.1-linux-amd64.tar.zst
sha256: f2d499917de99eadb856b104f51439544e15fc63a72d273489ecccf1bdc0d4fa
```

## Upstream attribution & licenses

| Artifact | Built from | License |
|---|---|---|
| `llama-server` b12110 | [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) at tag `b12110` | MIT (© The ggml authors) |
| `sd-server` 5dcfa79 | [leejet/stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp) at commit `5dcfa79` | MIT |
| `llama-moss-tts` / `llama-moss-tts-realtime` | [OpenBloom/llama.cpp](https://codeberg.org/OpenBloom/llama.cpp) at tag `openbloom-realtime-e1` (the OpenMOSS fork) | MIT |
| `tokenizer.json` | [OpenMOSS-Team/MOSS-TTS](https://huggingface.co/OpenMOSS-Team/MOSS-TTS) | Apache-2.0 |
| `MOSS-Audio-Tokenizer-ONNX` | [OpenMOSS-Team/MOSS-Audio-Tokenizer-ONNX](https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer-ONNX) | Apache-2.0 |

KrisKross is OpenBloom's *build* of components authored by the OpenMOSS team and contributors;
all credit for the upstream llama.cpp fork, the MOSS-TTS models and the MOSS Audio Tokenizer
belongs to them. The sidecars serve models published by their own authors under their own
licenses — the models are never hosted here.

## Integrity policy

- Every artifact in this repository is public and fetchable **unauthenticated** — this is a
  requirement of the OpenBloom downloader.
- Every artifact is content-addressed: the consuming installer pins a sha256 and fails closed
  on any drift.
- Never gate this repository, and never rewrite a published version in place — publish a new
  versioned path instead.