LocateAnything-3B Balance (448)
Balance is a separately compiled high-throughput model variant for RDK S600. It keeps the original LocateAnything-3B architecture and tokenizer, but uses a smaller Vision graph and a reduced Language graph configuration than Max (672).
Compilation Changes
| Compilation item | Max (672) | Balance (448) |
|---|---|---|
| Vision input canvas | 672 x 672 |
448 x 448 |
| Vision input shape | [1, 2304, 588] |
[1, 1024, 588] |
| Visual tokens | 576 | 256 |
| Prefill length | 1024 | 384 |
| KV cache | 4096 | 1024 |
| Maximum generated tokens | 4096 | 640 |
| Prefill graph | Standard | Fused Prefill |
| Logits graph | Standard | Compact Logits |
The 448 profile is calibrated and compiled with its own Vision and Language graph shapes. Its HBM outputs use the _balance suffix so they remain separate from the Max artifacts.
Model Files
Balance provides separately compiled Vision and Language HBM files. Keep the _balance suffix unchanged and use the matching Balance embedding file when one is provided. The tokenizer and model architecture remain compatible with the LocateAnything-3B model family.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support