Why this exists
- Comparesee what a model costs before pretending it is a free upgrade
- Documentturn one-off tests into reusable receipts
- Protectkeep the public site read-only so the GPU does not become community property
Research lane // live experiment
Local Model Lab
DriftLoom needed a place to compare local models without turning one internal GPU box into an open buffet for the internet. So this lab publishes bounded benchmark snapshots, model metadata, and operator notes instead of exposing a public prompt cannon.
live
Research
ollama
local-models
benchmarks
history-aware
v2 stance
The lab now keeps bounded run history instead of overwriting reality every time the runner wakes up. Same guarded public surface, better receipts, less goldfish-memory infrastructure.
Latest snapshot
Loading lab snapshot…
If this stalls, either JavaScript is sulking or the lab data is missing its receipts.
waiting on data
Host
- Status
- loading…
History
- Status
- loading…
Operator verdict
Loading verdict…
Model set
- Loadingthe model roster is still being assembled.
Immediate takeaways
Loading
Finding signal
Deriving rankings from the latest receipts.
Compact scoreboard
Balanced ranking
Working models, sorted for actual operator usefulness
Speed still matters, but so do footprint and whether the thing keeps tripping over reality.
Loading scoreboard…
Reading guide
How to read this without lying to yourself
Loading
…
Models under test
Prompt suite
Latest run results
Trend summary
Recent run history
Operator notes
- Loadingthe notes are still being dragged out of JSON.
Why the history matters
Local model performance is noisy. A single snapshot is a mood. A retained sequence is evidence.
Still deliberately bounded
No public prompts, no open benchmark API, no community GPU free-for-all. Read-only receipts are still the right shape for this box.