Laika MobileQAT Public Preview

Laika is a Gemma 4 E2B MobileQAT-derived local model artifact for the Laika macOS app, published as min-h-wei/Laika.

This repository is the public-preview model endpoint used by the signed Laika app. It contains the app-ready native MobileQAT bundle:

• model.safetensors: packed native MobileQAT weights
• config.json, tokenizer, processor, generation, and chat-template files
• laika_model_manifest.json: Laika runtime metadata
• ModelReleaseManifest.json: first-use download manifest with file sizes and SHA-256 hashes

The intended runtime is Laika's native Swift/Metal MobileQAT path (runtime_backend = mobile_qat_metal). This is not advertised as a generic drop-in transformers.generate() model; the checked-in config follows Gemma 4 metadata, but the packed weights and quantization layout are meant for Laika's native loader.

Preview Artifact

• HF repo: min-h-wei/Laika
• Laika manifest id: min-h-wei/Laika
• Laika model version: 2026.06.21-phaseF
• Runtime backend: mobile_qat_metal
• Source snapshot recorded by the bundle: 65707b8733090dda89f84735f1a1452e7b025f86
• Weight SHA-256: efab429012b97ab986c4d4838a46ff3ad95d618b42ce514771ca40fadc76a9a4
• Release manifest SHA-256: 48158f8a736b8e521a07293af84fdf6809bdecc43fc2469c4b197e36a63d05fe

This preview uses the current app-ready MobileQAT package already on main. It intentionally does not publish the newer June 30 diagnostic LoRA as the app model.

June 30 Diagnostic LoRA Status

The strongest June 30 training artifact is:

src/Model/artifacts/runs/20260630_runtime_residual_hard8_topwrong/04_personal_finance_patch_from_warm_ck48_iters24/lora_ck24.safetensors

That LoRA is a real diagnostic win:

• hard8+sentry replay: 175 / 175
• personal-finance full pack: 76 / 80
• independent probe: 39 / 39

It is not the promoted preview model. The training ledger marks it non-promotable because it is still a Python/native-bridge diagnostic LoRA, not a Swift-loadable native MobileQAT artifact with native parity, product gates, packaged-app smoke, and latency/RSS/swap gates.

Validation Snapshot

These numbers are local Laika gates. They describe this artifact in Laika's runtime path, not general benchmark performance.

Gate	Result
Native runtime smoke	pass
Technical-48	35 / 48 overall, answer 23 / 36, search 12 / 12
Product-shaped gates	personal raw 58 / 58 with 53 / 53 answers, synthetic base raw 12 / 12, synthetic extended raw 24 / 24, redaction 0 leaks
Held-out XPIA	27 / 27 raw, 27 / 27 citation, 27 / 27 grounding
Packaged light UI smoke	first answer about 17s UI-visible / 10.7s request; warm novel answer about 17.4s UI-visible / 10.8s request; exact cached answers under 120ms

These are preview validation numbers, not customer promises.

Limitations

• This is a Laika app-path artifact, not a full upstream Gemma release.
• The Technical-48 gate is not fully passing: the local report is 35 / 48.
• The preview is English-first and optimized for local personal files.
• Broad technical QA, multilingual search, exact count accuracy, OCR/photo breadth, and true agentic search are not headline claims for this build.
• Use requires compliance with Google's Gemma license terms.

Files

SHA-256 checksums for the app-critical bundle files:

efab429012b97ab986c4d4838a46ff3ad95d618b42ce514771ca40fadc76a9a4  model.safetensors
48158f8a736b8e521a07293af84fdf6809bdecc43fc2469c4b197e36a63d05fe  ModelReleaseManifest.json
cf6d7dc22738b5e6beb364bac833d78b869f5a6ffd57dfc96c6be3f2abc80424  config.json
cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f  tokenizer.json
26fed942f7b9112eff44c145a356d15a9f05b9a6e35e3917a1e70ebba4851f31  laika_model_manifest.json