CONCEPT Cited by 1 source
Cross-model portability¶
Definition¶
Cross-model portability is the property that switching the underlying model (or model provider) in a production system requires no caller-side code change — ideally only a configuration or model-name-string edit — while preserving the caller's existing prompt, parameter choice, and evaluation rubric.
It is the direct consequence of a unified parameter protocol applied across a model class (image generators, LLMs, embedding models, etc.).
Why it matters¶
Model portability changes the economics of model evaluation. Without portability:
- Each candidate model requires caller-side integration work.
- Teams settle on the first model that works, not the best one.
- Model-landscape churn forces repeated migrations.
With portability:
- Evaluating a new model is a config flip + re-run of the eval harness.
- Teams can A/B test model choices per-workload, per-project, per-A/B cohort.
- When a better model lands (new provider release, internal fine-tune), adoption is rollout-speed, not rewrite-speed.
The "best model varies by project" argument¶
From Instacart PIXEL (Source: sources/2025-07-17-instacart-introducing-pixel-instacarts-unified-image-generation-platform):
"An interesting outcome we realized from launching various applications was that the best performing model varied project by project. PIXEL enabled project leads to initiate projects using pre-configured, optimal model and parameter recommendations. Subsequently, they could rapidly test other models with a sample dataset and decide which one works best before moving to production image generation at scale."
The observation is canonical enough to anchor the architectural argument: there is no single best model for a portfolio of workloads, so portability is not a premature-optimisation luxury — it is how the portfolio gets good results at all.
Required mechanisms¶
Cross-model portability is load-bearing on three platform investments:
- Unified parameter protocol — canonical parameter vocabulary with platform- side translation.
- Single endpoint — one API surface that fronts all models.
- Evaluation harness / VLM-as-judge — model-agnostic quality evaluation that makes "is model A better than model B for this project?" quantitatively answerable.
The three compose: single endpoint routes the call, unified protocol translates parameters, evaluation harness scores the result. Any of the three missing and portability becomes nominal-only — technically possible but operationally expensive.
Seen in¶
- sources/2025-07-17-instacart-introducing-pixel-instacarts-unified-image-generation-platform
— canonical wiki instance at the image-generation layer.
PIXEL's
cfg_scale/style/sizenormalisation is the unified-parameter-protocol instance; the platform UI is the single-endpoint instance; the VLM-judge loop is the model-agnostic evaluation harness. All three together deliver the "switch by editing the model-name string" portability the post claims.
Related¶
- concepts/unified-parameter-protocol — required mechanism
- concepts/model-agnostic-ml-platform — the platform stance
- concepts/single-endpoint-abstraction — required mechanism
- patterns/ai-gateway-provider-abstraction — text-LLM sibling pattern
- patterns/unified-image-generation-platform — image-gen sibling pattern
- patterns/unified-inference-binding — SDK-level sibling pattern
- systems/instacart-pixel — canonical production instance
- companies/instacart