LayerCake is a research implementation of a modular language-model host. A LayerCake deployment combines an English core with independently packaged capabilities called cakes, then loads and executes only the capabilities a request selects.
Get started | Documentation | Architecture | Verified claims | Contributing
Project status: the scientific release is the immutable local tag
layercake-moonshot-finalat commit0537cbb9e93cd7ebd4ba01c0bf641414ecebb1c3. The default branch is a newer, unsealed host-development lineage. It does not inherit the release's Phase 2-8 certificates. See Project status before using benchmark numbers or making release claims.
Traditional model extension often couples every capability to one model checkpoint. LayerCake instead makes the host/package boundary explicit:
- cakes are immutable, content-addressed, non-executable archives;
- installation performs no receiver training or core mutation;
- package eligibility and automatic routing fail closed;
- inactive installed cakes do not receive proportional neural execution;
- request and response boundaries are UTF-8;
- persistent incremental state avoids recomputing completed context; and
- package identity, semantic behavior, and performance are measured as separate claims.
LayerCake hosts and executes capability artifacts. Extracting knowledge from a foreign teacher is a separate problem owned by the ABI project; ABI code and evidence are not vendored here.
| Goal | Starting point | Important boundary |
|---|---|---|
| Inspect, install, verify, and remove a local cake | Five-minute quickstart | Bundled example cakes are untrained lifecycle fixtures, not useful specialists. |
| Understand the system | Concepts and architecture | The sealed release and current development host are distinct lineages. |
| Integrate a compatible core and cake | Inference interface | You must supply artifacts matching the selected host interface. |
| Author a cake | Cake authoring | A new cake earns no quality or portability claim automatically. |
| Recompute the sealed campaign | Release verification | Use a detached exact-tag checkout and the required retained assets. |
| Navigate code, evidence, and historical experiments | Repository map | results/ is evidence; artifacts/ is model/package material; neither is the library API. |
LayerCake requires Python 3.10 or newer. From the repository root:
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -e .
python -m layercake --helpExercise the package manager with the bundled untrained Python fixture:
python -m layercake cake --registry .cache/demo-registry --catalog examples/catalog.json search python
python -m layercake cake --registry .cache/demo-registry install examples/python.cake --trusted-local
python -m layercake cake --registry .cache/demo-registry list
python -m layercake cake --registry .cache/demo-registry verify python
python -m layercake cake --registry .cache/demo-registry remove pythonThese commands demonstrate discovery, compatibility checking, content-addressed storage, integrity verification, and removal. They do not demonstrate model quality. Continue with the deployment quickstart when you have a compatible core and cake.
UTF-8 request
-> LayerCake execution host + persistent state
-> authenticated registry and package-bound routing policy
-> core-only, one selected cake, or an explicit orchestration plan
-> selected neural modules only
-> UTF-8 response + auditable execution trace
The primary implementation surfaces are:
layercake/cake/: package schema, signing, installation, and registry;layercake/models/: cores, portable decoders, and canonical interfaces;layercake/routing/: policies, routers, catalogs, and orchestration;layercake/runtime/: CPU, CUDA, native, and export paths;layercake/training/: research and fallback training workflows; andlayercake/evaluation/: quality, portability, performance, and campaign verification.
See the repository map before navigating the larger research and evidence surfaces.
The following are bounded results for the exact tagged release, artifacts, benchmark suite, comparator deployment, and laptop hardware named by the Phase 8 evidence. They are not universal performance claims.
| Area | Tagged-release result |
|---|---|
| Functional suite | LayerCake 100/100 on CPU and GPU; locked Qwen comparator 23/100 CPU and 24/100 GPU |
| CPU output-byte throughput ratio | 9.91x |
| GPU output-byte throughput ratio | 8.25x |
| LayerCake CPU vs comparator GPU | 7.67x |
| Held-out domain retention | 384/384 CPU, 384/384 GPU, identical on 384/384 |
| Routing checks | 1,980/1,980 |
| Hostile checks | 32 attacks across 24 categories |
The release does not establish universal language-model superiority, faster foundation training, physical mobile performance, calibrated energy dominance, arbitrary third-party host compatibility, latent multi-cake neural fusion, or external-laboratory independence. Read Verification and limits for the complete claim boundary.
LayerCake currently has two intentionally separate tracks:
- Sealed research release -
layercake-moonshot-final, the immutable evidence lineage used for the Phase 0-8 claims. - Post-release host development - the default branch, including newer signed direct-neural-core host constructs. These constructs prove mechanical host properties only until a successor campaign certifies a real artifact on the same lineage.
Phase 3 retired its original faster-training objective by governance. It is a sealed lifecycle disposition with zero training-efficiency claims, not proof of faster training and not a skipped phase.
Full-core training speed is a separate, currently open gate. LayerCake retains ordinary research and fallback training, but this release makes no training- dominance claim.
- New user: Deployment quickstart
- Integrator: Concepts, architecture, and registry specification
- Cake author: Authoring guide and threat model
- Research reviewer: Claims, phase status, and verification limits
- Contributor: Contributing guide and repository map
- ABI integrator: External artifact handoff
The numerous North Star, byte-model, training, and architecture-search files at the repository root are preserved research history. They are indexed in the repository map and are not the current product or release source of truth.
LayerCake is licensed under Apache-2.0. See LICENSE. Citation metadata is available in CITATION.cff.