Prism Engine

Prism Engine is a native compiler and runtime for inspectable, heterogeneous AI inference. It turns model weights and execution graphs into target-aware compute images (.cimage) and runs them through an ECS-native runtime with explicit placement, residency, scheduling, validation, and execution evidence.

Prism targets the machine that is actually available: Apple Silicon and Metal, portable CPU execution, AMD ROCm GPUs such as MI300X, and AMD XDNA/XDNA2 heterogeneous accelerator plans. The compiler can preserve mixed precision, progressively quantize or ternarize tensors, search KV-cache layouts, and route work across CPU, GPU, and NPU execution islands. Prism is pre-1.0 systems software, so APIs, artifact formats, and backend contracts remain subject to change.

What exists today

The important architectural distinction is that a backend result is not, by itself, authoritative. Prism’s compute kernel is being built so that state changes pass through validated transactions and durable receipts, while caches, queues, hardware handles, and analytical projections remain explicitly derived or ephemeral.

Repository map

For the governing architecture and its migration status, see compute-core/ARCHITECTURE.md. For the compute image ABI, see docs/cimage-layout-abi-v1.md.

Quick start

The most complete end-user path currently targets an Apple Silicon Mac with the Rust and Xcode command-line toolchains installed.

# Download a supported Hugging Face model and compile it to .cimage.
cargo run --release --bin prism --features full-apple -- \
  pull Qwen/Qwen2.5-0.5B-Instruct

# Start the local OpenAI-compatible server.
cargo run --release --bin prism --features full-apple -- \
  run qwen2.5-0.5b-instruct

# Send a chat completion.
curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"hello"}],"max_tokens":32}'

Compiled models are stored under ${PRISM_HOME:-$HOME/.prism}/models/<name>/. Each model directory contains its model.cimage alongside the configuration and tokenizer artifacts needed to load it.

The CLI also exposes prism list, prism compile <source>, and prism run <model> --chat. Run cargo run --bin prism -- --help for the current command surface.

Build profiles

Feature ownership differs between the root crate and tribunus-compute-core. Treat the two Cargo.toml files as the source of truth when embedding a specific surface. Off macOS, Metal kernels are replaced with a linkable placeholder. PRISM_MOCK_BUILD=1 forces that development mode on macOS; artifacts produced that way must not be shipped as accelerated builds.

Project status

Prism Engine is pre-1.0 research and systems software. Continuous integration checks the Apple Silicon prism-backend library and a narrower Linux CPU build. That does not imply that every declared binary, backend, model family, or experimental feature is production-ready. Supported behavior is the behavior covered by the selected build, test, hardware, and receipt gates.

The authority migration is also intentionally incremental. Several domains run in shadow mode while legacy registries are removed. The status in CAMPAIGN.md records that cutover work; “shadow” means the new transactional representation exists, not that it is already the sole runtime authority.

License

Prism Engine is available under a dual-license model: the GNU Affero General Public License, version 3 (AGPLv3), or a commercial license. The AGPLv3 option is the default for open-source use and carries its corresponding copyleft and network-use obligations. Organizations that need to distribute proprietary derivatives, keep modifications private, or operate without those AGPLv3 obligations can request a commercial license by emailing julian@tribunus.dev. See LICENSE for the AGPLv3 terms.