Gallery download
~75MB -> ~20MB4-bit storage-side variantsobserved4-bit Browser Gallery Models
This artifact is less glamorous than the model work, but it is exactly the kind of public detail users care about: what ships, how big it is, and what tradeoff it makes.
Headline Numbers
Evidence class applies to each value: measured = retained evaluation or runtime result; historical = recorded result not freshly reproduced; derived = calculated or decided from records; observed = current artifact state; not-measured = an explicit evidence gap.
Compression
~4xint4 gallery files vs fp16derivedBrowser-loadable
yesunlike Mac-side multi-GB specialistsobservedCompetitive Context
| System | Metric | Score | Size / Class | Comparable? | Readout |
|---|---|---|---|---|---|
| posttrainllm int4 gallery file | download size | ~20MB | 4-bit storage | Direct | Smaller cold-load distribution artifact for browser demos. |
| posttrainllm fp16 gallery file | download size | ~75MB | fp16 storage | Direct | Simpler baseline with larger download size. |
| Mac specialist packages | artifact size | multi-GB | HF/MLX safetensors | Directional | Different distribution lane; browser gallery compression should not be mixed with Mac specialist packaging. |
Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.
Distribution tradeoff
| Format | Benefit | Limit |
|---|---|---|
| fp16 | Simple and accurate | Larger download |
| int4 storage | Smaller cold load | Not a training-speed win |
Release Blockers
Not the active artifact channel
Browser gallery models are not the same as Mac-side specialist packages.
Unblock: Keep separate from specialist registry; only connect through public reports.