
Research/nonprofit watchAutomated record
Swiss AI Initiative
Capital-efficient model reach
01
PUBLIC TRACTION
Usage, audience and developer pull
701.6KHF model downloads29.2K per model
1.4KHF followers1.3K total likes
725GitHub stars246 forks
24Model portfolio30 datasets · 1 Spaces
WHY IT IS ON THE RADAR
Capital-efficient model reach
- High reach per model
- Shipping now
- GitHub momentum
02
MODEL FOOTPRINT
What it publishes
Language/TextMultimodalRetrieval/Embeddings
Most popular HF repositoryswiss-ai/Apertus-8B-Instruct-2509623.8K downloads · 489 likes ↗
- Top tasks
- image-text-to-text (2), text-generation (20), feature-extraction (1)
- Model sizes
- <0.5B (8); 0.5-3B (7); 3-10B (5); 30-100B (3); Unknown (1)
- Base families
- BERT
- Libraries
- transformers
03
DEVELOPER ACTIVITY
How it ships
Original repositories79
Forked repositories46
Pushed in 90 days58
Archived share5%
Most-starred repositoryswiss-ai/pretrain-data154 stars · 34 forks ↗
- Languages
- Python, Jupyter Notebook, Shell, Dockerfile, HTML, TypeScript, Go
- Topics
- apertus, hardware-information, hpc-clusters, performance-monitoring, open-source-ai, whitepaper, data-pipeline, dataset, jats, open-access
- Latest release
- swiss-ai/serving-api · Aug 20, 2026
04
ACTIVITY TIMELINE
Latest observed moves
- GitHub pushRepository activity ↗
- HF datasetDataset published or updated ↗
- GitHub releaseRelease v0.0.34 ↗
- HF modelModel published or updated ↗
- HF SpaceSpace published or updated ↗
05
THE NEOLAB STACK
Publicly evidenced tooling
- Training
- MLX, Megatron-LM, Transformers
- Compute
- Not disclosed
- Hardware
- Not disclosed
- Tracking
- Not disclosed
- Bottlenecks
- Licensing / rights, Evaluation
“Not disclosed” means no evidence was found in the sampled public metadata or top model card—not that the lab does not use the tool.
06
COMPARABLE LABS
Nearest modality peers
INVESTOR DILIGENCE
Confirm the team and compute inputs behind reach per model; benchmark quality versus the base family.
- HF member count is only a public-profile proxy
MISSING FROM THIS RECORD
foundedYearfoundersOrLeadershipemployeeCountfundingbusinessModelproducts
Completeness: 8 of 14 tracked field groups. Unknowns are explicit so an analyst can close them.
DATA PROVENANCE
Every claim should be auditable.
Platform counters are a point-in-time public snapshot. Editorial facts are only promoted after source review.
NEOLAB TRACKER PRO
Know when the thesis changes.
Weekly deltas, funding and team alerts, portfolio watchlists, CSV exports and an analyst-ready API turn a static profile into a sourcing system.
TensorBlock
SupraLabs
Cactus Compute, Inc.
NuMind
VAGO solutions