Primary use stated or inferred cautiously from official documentation.
B
BYTEDANCE SEED / ByteDance-Seed/Qwen3-30B-A3B-MoE-SAE
Qwen3-30B-A3B MoE-Output SAE
Three 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.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | ByteDance-Seed/Qwen3-30B-A3B-MoE-SAE |
|---|---|
| Model type | Interpretability |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 05 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Internal activations |
| Exit | Latent features · Reconstructed activations |
| Context window | Qwen3-30B-A3B layers 4, 30 and 46 |
| maximum output | 100 active features per token per autoencoder |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · PyTorch |
Study which experts contribute to particular features and quantify activation reconstruction error. The three layer checkpoints share one profile. They do not replace the base model, generate text or serve as document embeddings.
Each FP32 weight file takes about 17.2 GB, on top of base-model memory. Layers do not share an interchangeable latent space. Long inputs may need chunking. Observed attribution alone does not establish causality.
03 / CAPABILITIES
What it can do
Orientation
Study which experts contribute to particular features and quantify activation reconstruction error. The three layer checkpoints share one profile. They do not replace the base model, generate text or serve as document embeddings.
Context
Qwen3-30B-A3B layers 4, 30 and 46 · 100 active features per token per autoencoder
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
Internal activations
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No hosted commercial API has been verified for this checkpoint. Downloadable weights do not cover storage, accelerators, memory or operating costs.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
ByteDance reports 12.6%, 25.1% and 7.6% unexplained variance for layers 4, 30 and 46 respectively. Lower means better reconstruction in that test. These are not task accuracy rates or chat model rankings. Inferama has not reproduced the measurements.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| Qwen3 MoE SAE FVU (layer 4) | 12,6 % | Lower is better | View source ↗ |
| Qwen3 MoE SAE FVU (layer 30) | 25,1 % | Lower is better | View source ↗ |
| Qwen3 MoE SAE FVU (layer 46) | 7,6 % | Lower is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Qwen3-30B-A3B MoE-Output SAE
Version added to Inferama's verified catalog.
07 / SOURCES