ModelSpec is an analysis of alternatives for AI models: it decides which models fit a job from your requirements, sourced benchmarks and real cost, and shows its work.
Retrieval embeddings, self-hosted
Find a commercially usable open-weights embedding model that you can run yourself.
Worked example as of 2026-09-23, from the public data image.
Open weights you run yourself; commercial licence.
Requirements
model.class = vectoriser: Retrieval embeddings require the model class that emits vectors.model.weights_openness = open_weights: Self-hosting requires downloadable weights; add a model.fits_hardware Must for the device you own.licence.commercial_use in {permitted, permitted_with_conditions}: The licence must allow commercial use, including use with stated conditions.
Preferences
retrieval: Rank only by retrieval quality among self-hostable embedders.
Applied template spec
spec_version- 1
where- model.class = vectoriser
- model.weights_openness = open_weights
- licence.commercial_use in {permitted, permitted_with_conditions}
optimizeweightsretrieval- 1.0
Worked answer
Decision status: answered.
Separated on the evidence
kingsoft/qzhou-embedding
probability bands: a model joins the leader when P(its score >= the leader's) >= 0.25 under the score posteriors; a weighted capability with an 80% interval wider than 2.8 is not enough evidence
probability bands: a model joins the leader when P(its score >= the leader's) >= 0.25 under the score posteriors; a weighted capability with an 80% interval wider than 2.8 is not enough evidence
Score and score interval use the objective's dimensionless normalised units. Cost per task is for the engine-selected offering and the applied spec's task size. When task size is omitted, the engine uses its documented default. P(best) is the engine probability. Unavailable means the engine did not return a value. Thin evidence is undetermined and is not a recommendation.
Best band
| Model | Score | Score interval | P(best) | Cost per task, USD | Capability estimates and intervals |
|---|---|---|---|---|---|
kingsoft/qzhou-embedding | 0.0 | [-0.910881, 0.910881] | 0.10546875 | Unavailable |
|
Thin evidence: undetermined
| Model | Score | Score interval | P(best) | Cost per task, USD | Capability estimates and intervals |
|---|---|---|---|---|---|
microsoft/harrier-oss-v1-27b | 1.0 | [-0.872477, 2.872477] | 0.51953125 | Unavailable |
|
tencent/kalm-embedding-gemma3-12b-2511 | 0.600425 | [-1.272051, 2.472902] | 0.375 | Unavailable |
|
Engine warnings
- not_separable
Decision dec_fb3791921baf232164b49490 on public snapshot snap_05a963d2477ac2fb.
What this template teaches
Privacy is a Must on the weights and the licence; the canvas shows the model size you would have to host. Lab jurisdiction is a buyer-set Must (origin.lab_jurisdiction in {…}); an unverified jurisdiction does not pass, so set the countries on the facet board. The same condition works as an inventory profile rule.
The live board uses current data. Its answer can change as evidence changes.
Human lookup is free on Decide. CLI, API, and MCP access requires a key. Access and pricing.