BASE-1
A multimodal foundation model whose architecture is discovered through decentralized neural architecture search
Status: In Development
BASE-1 is currently under active development. No weights are available yet. This repository will host the model checkpoints, configuration, and usage documentation once the architecture search and training phases are complete.
Model Summary
| Developer | Cortex Foundation in partnership with Platform |
| Architecture | Determined by neural architecture search (in progress) |
| Parameters | To be announced after architecture search |
| Input modalities | Text, Image |
| Output modality | Text |
| Architecture search | PRISM — decentralized NAS on the Platform Network |
| License | Apache 2.0 |
Overview
BASE-1 is a foundation model being developed through PRISM, a decentralized neural architecture search (NAS) challenge running on the Platform Network. Rather than committing to a hand-designed architecture upfront, BASE-1's design is discovered competitively: miners across the network submit novel architecture families and training recipes, which are evaluated in isolated benchmark environments for learning quality, training stability, and scaling behavior.
The best-performing architecture that emerges from this search will be used to train BASE-1 at scale.
How the architecture is discovered
PRISM fixes the dataset and evaluation protocol, not the search space. Candidate submissions are scored on:
- Learning quality — proxy loss performance under a shared, deterministic evaluation contract
- Training stability — smooth loss curves, stable gradients, and well-behaved activations
- Scaling signals — consistent improvements across model size, depth, sequence length, and batch scaling
- Noise resistance — dynamic thresholds prevent marginal random fluctuations from being rewarded as improvements
Architecture discovery and training-recipe improvements (optimizer, loss computation, inference, train step) are attributed and rewarded independently, so both the model design and its training procedure are optimized by the network.
Modalities
BASE-1 will support Text/Image to Text: it will accept text and images as input and generate text as output.
| Input | Output |
|---|---|
| Text | Text |
| Image | Text |
Why is the model size not announced?
The parameter count of BASE-1 is genuinely not decided yet — and this is by design.
In a conventional training pipeline, the architecture and parameter budget are fixed first, then training begins. BASE-1 inverts this process:
- Architecture search comes first. PRISM evaluates candidate architectures at compact proxy scales, measuring loss curves, gradient stability, activation behavior, and how performance evolves across model size, depth, sequence length, and batch size.
- Scaling laws are derived from the winning architecture. Each architecture family exhibits its own scaling behavior. The optimal parameter count depends on the scaling-law signals of the architecture that wins the search — a number that cannot be known before the search concludes.
- The final size is chosen from evidence, not convention. Once the winning architecture's scaling characteristics are measured, the parameter budget will be set where the compute/performance trade-off is optimal for that specific design.
The final model size will be announced once the architecture search is complete.
Roadmap
| Phase | Description | Status |
|---|---|---|
| 1. PRISM challenge launch | Open the decentralized architecture search to miners on the Platform Network | In progress |
| 2. Architecture selection | Identify the best-performing architecture family from competitive evaluation and scaling analysis | Pending |
| 3. Dataset curation | Assemble and validate the large-scale multimodal training corpus | Pending |
| 4. Large-scale training | Train BASE-1 at the parameter budget derived from the winning architecture's scaling laws | Pending |
| 5. Model release | Publish weights, configuration, evaluation results, and usage documentation in this repository | Pending |
Intended Use
BASE-1 is intended as a general-purpose multimodal foundation model for text generation conditioned on text and image inputs. Detailed intended-use guidance, limitations, and evaluation results will be published with the model release.
Evaluation
Benchmark results will be published alongside the weights once training is complete. Architecture-search-stage evaluations follow the PRISM scoring protocol, documented in Scoring and rewards and Scaling evaluation.
Resources
- PRISM (architecture search): github.com/PlatformNetwork/prism
- PRISM documentation: Overview | Architecture | Scoring | Scaling
- Platform Network: platform.network
License
This repository is released under the Apache License 2.0.