Eligible on declared metadata
Model role, source classification and arithmetic weight-only budget are consistent with the chosen constraints. This does not imply agent task success.
AGENT INTELLIGENCE · SOURCE-FIRST
Agent Model Selector
Narrow the verified Leading and Notable model selection with traceable constraints and source-level caveats, without fictional scores.
CONSTRAINT-AWARE MODEL DISCOVERY
Choose requirements, not a ranking.
The decision support below runs entirely in your browser. It distinguishes hard metadata filters from evidence gaps. A matching record is a candidate to evaluate, not a certified system.
Choose your constraints to inspect the 31 catalog records.
How to interpret the results
Needs verification
Capability, runtime, permissions or hardware are untested or not sufficiently documented. Unknown does not count as supported.
Excluded by explicit constraints
When a known classification conflicts with a hard filter, the record is excluded from the displayed shortlist; nothing is called a failed benchmark.
Local-first choices
Model files alone cannot guarantee a local-only system. Check embeddings, tools, logging, backups and network requests.
DIRECT ANSWERS
Questions about this evidence
Does the selector recommend the best agent model?
No. It identifies candidates compatible with selected catalog criteria and explicitly lists unresolved requirements. It has no quality score or proprietary performance ranking.
Is publisher-documented tool use independently verified?
No. It is a source-backed vendor claim. Verify the exact checkpoint, runtime and task on your own harness before deployment.
Does an estimated weight footprint guarantee VRAM fit?
No. Mathematical BF16 weight size omits KV cache, activations and serving overhead; runtime observations are needed.
Go from evidence to decisions.
Documented interfaces and publisher-run benchmarks do not certify autonomous production deployments.
Open the constraint-aware selector → · Read our methodology →