Free2AITools Nexus Index (FNI)

The FNI is the ranking algorithm behind Free2AITools. It evaluates the full catalog of AI entities (models, datasets, papers, tools, agents, spaces, prompts) across five weighted factors to produce a single 0-100 score.

Current version: V2.0

The Formula: S.A.P.R.Q

FNI = 0.35*S + 0.25*A + 0.15*P + 0.15*R + 0.10*Q

Factor	Weight	What It Measures
S — Semantic	35%	Query-time relevance (currently dormant: keyword-index based; live semantic/ANN ranking is not currently provided; reported as null + a note on static surfaces)
A — Authority	25%	Ecosystem gravity: knowledge mesh centrality, cross-entity citations, source credibility
P — Popularity	15%	Community adoption: downloads, stars, likes (log-scaled to prevent gaming)
R — Recency	15%	Freshness: exponential time decay with type-specific half-lives
Q — Quality	10%	Completeness: README depth, metadata richness, runtime compatibility

Agent Structured Tags (V2.0)

Beyond the five score factors, FNI V2.0 attaches structured metadata for machine-readable catalog filtering:

Tag	Type	Meaning
ollama_compatible	boolean	Has GGUF quantization files — can be run via ollama run
can_run_local	boolean	Locally runnable: ≤13B parameters + GGUF available
license_type	string	Classified as permissive, copyleft, non-commercial, or unknown
hosted_on	string[]	Cloud providers offering this model (Replicate, Together, HF Inference)
hosted_on_checked_at	ISO date	When the hosting data was last verified

These tags power the select_model API — AI agents can filter the catalog by these stored-metadata fields (size, license type, GGUF/Ollama indicators) via MCP or HTTP. They are heuristic metadata filters, not verified runtime compatibility checks.

Interpreting FNI Scores

Range	Label	Meaning
80-100	Very high FNI signal	Strong across all five factors
60-79	High FNI signal	Well-rounded with clear strengths in 3+ factors
40-59	Medium FNI signal	Strong in 1-2 factors, mid-range elsewhere
20-39	Low FNI signal	New or niche — may be rising fast (check R factor)
0-19	Insufficient FNI signal	Minimal community footprint or very stale

Version History

V1.0 (2025) — Originally called “Freshness-Novelty Index.” Four-factor P.V.C.U formula (Popularity, Velocity, Context, Uniqueness). Single-source HuggingFace data. No semantic component.

V2.0 (2026) — Renamed to “Free2AITools Nexus Index.” Five-factor S.A.P.R.Q formula. Multi-source aggregation across 7+ platforms. Catalog re-evaluated daily. Agent structured tags for select_model API. Factor scores included in every API response.

Source-Parity Weighting (P factor)

The Popularity factor compares raw engagement (downloads, stars, likes, citations) across platforms whose metrics are not directly comparable — an arXiv citation and a HuggingFace download are not the same unit of signal. To put them on one scale, raw popularity is multiplied by a per-source coefficient (Ks) before log-compression. These coefficients are an editorial weighting that favors peer-reviewed academic provenance, and they are published here so the weighting is transparent and contestable:

Source	Coefficient (Ks)	Rationale
arXiv	30.0	Knowledge roots — peer-reviewed / academic provenance
Semantic Scholar	30.0	Knowledge roots — peer-reviewed / academic provenance
GitHub	5.0	Tool source — engineering adoption signal
HuggingFace	1.0	Model forge — baseline (engagement is high-volume, low-cost)
CivitAI / other	0.2	Community market — high-volume, lower verification

These are the live values in scripts/factory/lib/fni-score.js (SOURCE_COEFFICIENTS). The coefficient is applied as raw_popularity × Ks, then asymptotically log-compressed (base 8) into the 0-100 P factor, so a high coefficient lifts a source’s floor without letting any single metric dominate.

Anti-Gaming

FNI uses multi-dimensional cross-validation to detect manipulation:

• Anomalous growth: alerts when 7-day metric growth exceeds 10x the category average
• Ratio anomalies: download/star ratios outside reasonable ranges for the entity type
• Content mismatch: high popularity but no substantial documentation or code

Flagged entities are reviewed and scores adjusted. Log-scaling on popularity metrics (P factor) inherently dampens artificial inflation.

For Developers

Every API response includes fni_version and per-factor breakdown. The Semantic factor is scored live at query time, so on static detail/select/compare surfaces it is reported as not-measured (null + a note) rather than a placeholder value — honest-contract: a constant is not a measurement:

{
  "fni_score": 72.4,
  "fni_factors": {
    "semantic": null,
    "semantic_note": "query-time baseline; scored live at search; not a per-entity value",
    "authority": 85.2,
    "popularity": 67.1,
    "recency": 91.0,
    "quality": 44.3
  }
}

The FNI algorithm is open source: github.com/mosesy5688-cell/ai-nexus

Related

• Methodology — Visual breakdown of the FNI formula and fairness pillars
• Developers — API documentation for select_model and compare
• LLM Benchmarks — Traditional evaluation metrics (MMLU, HumanEval)
• Local Inference — Running models locally with Ollama/llama.cpp
• VRAM Requirements — Hardware needed for different model sizes