lokai-mlf-catalog

Lokai MLF Runtime Pack Catalog

Status: public model-distribution repo; the Lokai engine and public stack are coming soon. Generated: 2026-10-28T20:57:11Z

This repository is the model-distribution lane for Lokai MLF packs. The installable list contains only packs that passed the current modular-runtime, provenance, license, installation, and behavior gates. Separately labeled candidate artifacts may be present for qualification; their presence does not make them installable.

Lokai model packs are installed by the Lokai Sovereign Public Stack. They are not standalone Transformers, GGUF, safetensors, PyTorch, or llama.cpp releases. Each pack combines MLF payloads with public-safe metadata for Lokai's modular native runtime.

Start Here

1. Choose a model from Available Modular Packs below.
2. Select its Browse files link to see the downloadable archive, manifest, size, checksum, source, and license evidence.
3. At launch, install the pack from Lokai's Models screen. Lokai handles the download, verification, device compatibility check, and activation.

The model archives are visible under Files and versions. They are runtime packs, not standalone Transformers or llama.cpp models, so downloading one now is useful for inspection but requires the Lokai public stack for execution.

What MLF Is

MLF is Lokai's shard-addressable container format. Today it packages model tensors, tensor maps, tokenizer/configuration material, checksums, provenance, and runtime-binding metadata so Lokai can validate and move only the data needed by the selected execution path instead of treating a model as one opaque file.

The container contract is intentionally broader than language-model weights. Typed MLF shards can represent independently verifiable pieces of multimodal models and, as those public runtimes qualify, video, image, and audio data. The same signed, indexed container design is also intended for composable software and operating-system assets. Those future data types are roadmap capabilities, not a claim that every payload is executable in the current preview.

MLF does not remove upstream licenses or provenance. Every published pack keeps its source identity, terms, byte size, and content hashes, and the Lokai runtime must verify those boundaries before activation.

Available Modular Packs

Use each row's Browse files link to inspect its archive and manifest in Hugging Face's Files and versions view. Normal users should install packs from Lokai's Models screen rather than downloading and unpacking files manually.

Only packs proven through the modular runtime and product installation gates are published as available. Additional model and hardware support will appear only after passing the same provenance, execution, behavior, and clean-install gates.

Public Candidate Packs

These artifacts are approved for anonymous redistribution but remain disabled in the product installer until the stated gate passes.

Presence in this table is not runtime promotion, a default-download claim, or public-GA approval.

How Distribution Works

• GitHub will ship the Lokai Sovereign Public Stack: product source, web control surface, setup and recovery flows, public runtime contracts, validators, LAN Swarm tools, public-safe seed memory, and extension surfaces.
• Hugging Face hosts model catalogs, source registries, signed/checksummed pack manifests, license/provenance evidence, and optional model-pack artifacts.
• Model weights are optional runtime packs. They are not bundled into the GitHub source release.
• Users select and install packs from Lokai's Models panel. The installer records terms acceptance, downloads resumably, verifies size and SHA-256, rejects unsafe archive paths, validates the modular pack, and only then binds it to the local Lokai runtime.
• Models remain optional. A source checkout never bundles model weights.
• Torrent mirrors may be added later for public packs that pass the same gates.

Required Runtime

• Product/runtime: Lokai Sovereign Public Stack
• Public stack URL: https://github.com/Lokai-io/lokai
• Source status: this model catalog is public; the Lokai engine and installable public stack are coming soon at https://github.com/Lokai-io/lokai
• Engine path: modular native Lokai runtime selected through hardware-aware capability binding
• Expected install root for packs: public_runtime/
• Expected weight root inside a staged pack: public_runtime/weights/packs/active/<model_id>/
• Standalone raw-weight use: not supported by this repo

Downloading a pack without the Lokai public stack does not create a usable Lokai installation. The product supplies validation, device selection, grounding, completion controls, tools, governance, and the deployment-local runtime binding.

First-Use Experience

1. Watch https://github.com/Lokai-io/lokai for the engine and public-stack release. The model packs can be inspected now, but they require Lokai for governed execution.
2. Open Models. Lokai recommends only packs compatible with the detected device.
3. Review the model size, upstream source, license, and intended use, then approve the download.
4. Leave the page open or continue elsewhere. Download and verification progress remain visible and resumable.
5. When validation passes, open Chat or a project. Lokai selects grounding, tools, runtime settings, and completion policy automatically; advanced users can inspect or override eligible controls.

No Hugging Face account or token is required for public-GA packs.

What A Model Pack Contains

Approved pack artifacts are zip archives whose root is public_runtime/. A pack normally contains:

• sanitized model config, sidecar, tensor map, tokenizer, and shard index files
• Lokai MLF weight shards under public_runtime/weights/packs/active/<model_id>/
• a pack contents manifest with per-file SHA-256 checksums
• a pack manifest with archive checksum, size, upstream model reference, license evidence, and runtime requirement metadata

The repository card uses license: other because every pack retains its own upstream terms. The license and provenance fields in each pack manifest are the authoritative terms for that artifact.

Pack manifests are designed so the Lokai installer can verify the artifact before making a model selectable.

Repository Contents

• README.md
• public-safe catalog metadata, for example catalog/trusted_model_catalog.json
• public-safe source registry metadata, for example catalog/model_source_registry.json
• signed/checksummed pack manifests
• checksums
• license/provenance files
• optional Lokai MLF model-pack artifacts only when redistribution is allowed by the upstream license and an explicit release gate approves them

This catalog contains reviewed distribution metadata and approved model-pack artifacts. Model artifacts appear only when redistribution is allowed and the publication workflow binds them to provenance, license, manifest, size, and checksum evidence.

Extending Lokai

The public stack is intended to be extendable through stable, versioned contracts. Model and backend contributors work against public contracts for manifests, tensor maps, runtime modules, hardware capabilities, receipts, and fail-closed validation. A contribution does not become a trusted public pack merely because it runs locally: it must preserve provenance, pass architecture and capability tests, produce auditable receipts, and complete the community review and maintainer promotion path.

The initial catalog is intentionally narrow so installation remains truthful and repeatable. Contributor contracts and verification guidance will be published with the Lokai public stack at https://github.com/Lokai-io/lokai.