Laboratory
Entity that develops or publishes the model and maintains its technical and safety documentation.
Explore the documented models, published evaluations and primary sources associated with this organisation.
ByteDance Seed sparse transcoder for studying Qwen3-8B-Base. It reads the residual stream before the MLPs and approximates their outputs using features across layers. The verified revision contains all 36 encoders and 36 decoders; uploading finished on November 5.
06 NOV 2026↗ PER-LAYER SPARSE AUTOENCODERSThree ByteDance Seed TopK autoencoders analyze Qwen3-30B-A3B MoE block outputs. The weights and research paper are dated November 5. Each layer has a separate checkpoint with 1,048,576 latent features, selecting up to 100 per token.
06 NOV 2026↗Figures are shown with the context reported by their source. A provider result is not an independent comparison and does not replace your own evaluation.
| Model | Benchmark | Result | Metric |
|---|---|---|---|
| Qwen3-30B-A3B MoE-Output SAE | Qwen3 MoE SAE FVU (layer 4) | 12,6 % | Lower is better |
| Qwen3-30B-A3B MoE-Output SAE | Qwen3 MoE SAE FVU (layer 30) | 25,1 % | Lower is better |
| Qwen3-30B-A3B MoE-Output SAE | Qwen3 MoE SAE FVU (layer 46) | 7,6 % | Lower is better |
An organization, a product, and a model are not the same unit. Inferama separates them to avoid attributing capabilities or commercial terms to the wrong item.
Entity that develops or publishes the model and maintains its technical and safety documentation.
Identifiable version with limits, modalities, and behavior that may change between releases.
API, application, associated cloud, or commercial plan; each channel may have different pricing, retention, and limits.
Every claim must retain the official page consulted and the verification date.