What this model does
Qwen3-30B-A3B MoE-Output SAE is not a conversational model or a version of Qwen that generates responses. It is a set of three sparse autoencoders designed to analyze internal outputs from mixture-of-experts (MoE) blocks in Qwen3-30B-A3B. Each autoencoder corresponds to a specific layer: 4, 30, or 46. The package was released alongside a research paper submitted to arXiv on November 5, 2026.
An MoE block combines the outputs of several experts selected by a routing mechanism. The SAE transforms that output into activations of latent features and can reconstruct it from those activations. In this case, each layer has a feature space of 1,048,576 latents, and the TopK activation retains up to 100 per token. Features from different layers do not share an interchangeable space: a feature in layer 30 should not be treated as if it had the same meaning or function as one in layer 4.
The paper asks how experts’ contributions to latent features are distributed, not what answer Qwen would produce in response to a prompt. The authors argue that decomposing activations makes it possible to analyze specializations more specific than broad subject categories. The findings are limited to the model, layers, data, and method studied; they do not establish a general rule for all MoE models.
How it extracts features and attributes contributions
The SAE takes the MoE block output without normalization, before the residual addition. Its encoder applies an affine transformation, a ReLU function, and TopK selection; the decoder combines the selected features to reconstruct the output. TopK limits how many features are retained per token, but it does not turn their interpretation into a deterministic classification of what the model “understands.”
The study’s central idea uses the fact that the block output is a weighted sum of expert outputs and that the encoder is affine. The authors therefore decompose each feature’s pre-activation across the experts that contributed to the output. Aggregating those contributions over a corpus produces sets of features that each expert primarily drives. This calculation estimates an association under the defined procedure; it does not show that an expert alone causes a capability or that a feature has one unique semantic interpretation.
The autoencoders were trained separately on activations collected from the Olmo 3 training corpus: five-eighths Dolma 3 pretraining data, one-quarter Dolci-Think-SFT, and one-eighth Dolci-Instruct-SFT. The model card reports approximately 3.2 billion tokens per SAE. The repository also describes a validation set of about 1.5 billion tokens, separate from the SAE training data and drawn from a similar mix of sources. The composition and separation of these datasets bound the claims the analysis can support; they are not equivalent to an evaluation across all possible uses or languages.
The project then applies a stage of feature auto-interpretation and expert labeling. The repository states that GPT-5.5 generated the feature explanations, expert labels, thematic groups, and certain judgments. These are automated outputs, not human annotations. A feature explanation can help explore patterns, but it should be treated as an interpretive hypothesis rather than a verified account of the model’s internal function.
From an activation to a hypothesis about specialization
- 01Capture the output of a specific Qwen3-30B-A3B MoE layer.
- 02The SAE encodes that output and retains up to 100 positive activations per token.
- 03The method decomposes each feature’s pre-activation across experts and aggregates their contributions over the validation corpus.
- 04Dominated features are grouped to study patterns, and automated explanations are generated; these can guide analysis but do not verify causality.
What results the study reports
The preprint reports that, in layers 4, 30, and 46 of Qwen3-30B-A3B, contributions to features are more concentrated than expert selection as measured by routing. It also says that most analyzed features have one or two dominant experts and that overlap between dominant sets is concentrated in a few expert pairs. These are findings for the study’s specific setup, not an independent evaluation of the model or proof that the same pattern appears in other systems.
Examples in the paper include divisions of labor more fine-grained than a broad “programming” label: different experts are associated with Java and front-end web code. In mathematics, the analysis distinguishes experts linked to geometry from those linked to arithmetic and polynomials. These examples illustrate the authors’ interpretations; they do not mean that each expert is limited to one task or that the model has a separate module for every domain.
The reconstruction experiment reports an unexplained fraction of variance (FVU) of 12.6% for layer 4, 25.1% for layer 30, and 7.6% for layer 46. A lower FVU means the SAE leaves less variance unexplained in that experiment. FVU does not measure task accuracy, response quality, safety, or usefulness to end users. The model card attributes these results to the paper; they are not an independent reproduction.
The available paper is arXiv preprint version v1, submitted on November 5, 2026. The record does not establish that it has undergone peer review. The team’s code repository provides experiments that can be run using the settings described in the paper, along with derived data and some explanations; this makes the procedure easier to inspect but is not equivalent to independent replication. Exact reproduction of the automated explanations is also not guaranteed: the repository notes that new samples and explanations may produce different results.
How to interpret the metrics and findings
| Published result | What it supports | What it does not establish |
|---|---|---|
| FVU: 12.6% in layer 4, 25.1% in layer 30, and 7.6% in layer 46 | The SAE reconstructs outputs with different levels of unexplained variance in the described experiment. | It is not an accuracy rate, reasoning score, or overall model comparison. |
| Feature contributions appear more concentrated than routing | In the studied corpus and layers, the two ways of measuring specialization show different distributions. | It does not prove that the observed contribution causes an ability or that the pattern is universal. |
| Examples of experts associated with Java, web code, geometry, or arithmetic | The labels illustrate the level of granularity the analysis attributes to certain experts. | They do not prove exclusive subject areas or human validation of every label. |
Cost of use and practical requirements
Each SAE’s weights occupy approximately 17.2 GB in float32, in addition to the memory needed to load the base model. The model card says the dense pre-activation matrix requires 4 MiB per token; it recommends processing long inputs in chunks. The reproduction code reports that Qwen3-30B-A3B in bfloat16 uses about 61 GB and that the SAE encoder adds about 8 GB for each traced layer. According to the team’s documented setup, tracing all three layers in one pass requires two 80 GB GPUs.
No official API price has been verified for this checkpoint, and the model page does not indicate deployment by an inference provider. Downloading the weights does not make the analysis free: costs depend on storage, accelerators, memory, and operating your own environment. The resource figures describe the repository’s procedure; actual infrastructure needs can vary with hardware, precision, and execution method.
For a simple experiment, the repository shows how to load the weights, capture a layer’s output, and encode it. The example uses layer 30 and a short input; it should not be mistaken for a speed benchmark or a universal production recipe. Compatibility depends in particular on the Qwen3 interfaces in Transformers and on capturing the signal expected by the SAE.
The code repository says that collecting the statistics requires a Transformers version earlier than 5, and that experiments based on language models use an OpenAI-compatible endpoint. Reproducing the full analysis therefore takes more than downloading the three files: it requires prepared validation data, sufficient memory, suitable dependencies, and access to a compatible service for certain interpretation stages. Publicly available weights do not remove those dependencies.
Choosing based on your objective
| Objective | Is this SAE a fit? | Main condition or limitation |
|---|---|---|
| Inspect how feature contributions are distributed in the studied layers | Yes, as an analysis tool for the exact Qwen3-30B-A3B model. | Use the matching layer and follow the published activation procedure. |
| Use a chatbot or deploy text generation through an API | Not on its own. | The SAE does not generate conversations, and no commercial API was verified for this checkpoint. |
| Claim that an expert causes a specific ability | Not as an automatic conclusion. | Contributions and labels do not constitute causal evidence. |
| Repeat the run with long inputs | Potentially, by processing inputs in chunks. | Memory costs are substantial, and dense pre-activations consume resources per token. |
Limitations, safety, and licenses
A latent feature is a learned direction for representing patterns in activations; its name or natural-language explanation is a subsequent interpretation. A plausible explanation does not prove that a feature has a unique, stable, or complete meaning. In this project, automated interpretations and GPT-5.5-generated labels should be distinguished from the mathematical measurements calculated from activations.
The proposed decomposition examines contributions within a particular expert mixture and encoder. It is not a safety audit, robustness test, or runtime control mechanism. Nor does identifying a feature guarantee that modifying it will predictably change the model’s behavior. An internal association is not a causal intervention.
The measurements are limited to three layers of a specific Qwen3-30B-A3B version and to the validation distribution described by the team. The published detailed explanations cover only certain experts in layer 30; the repository says that some results require generating additional explanations and that different samples will not necessarily reproduce the paper’s tables or labels exactly. These interpretations should not be extrapolated to every layer, model, task, or language.
The model card states that the SAE and base model are released under the Apache License 2.0. The activation data used for training come from Dolma 3, Dolci-Think-SFT, and Dolci-Instruct-SFT, whose license is identified as ODC-BY v1.0. The repository also says it does not redistribute the original corpora and that derived materials remain subject to the relevant conditions; anyone reusing weights, code, or data should check the terms that apply to each component.
When to choose it and how to evaluate it
This artifact fits researchers or technical teams who want to study the internal organization of Qwen3-30B-A3B and can afford to run the model and SAEs. It is less suitable if the goal is to compare chat quality, measure an end-user task, or purchase hosted inference: those questions require different evaluations and services, which this set of weights does not provide.
For your own evaluation, first define the claim you want to test: activation reconstruction, concentration of contributions, stability of an interpretation, or the usefulness of an intervention procedure. Then keep the metrics separate. FVU addresses reconstruction; contribution analysis addresses how a calculated quantity is distributed; an automated label proposes an interpretation. None substitutes for the others.
A comparison between layers should preserve each SAE’s context and use the same capture procedure. To test stability, you can vary validation documents and feature samples, recording automated explanations and contribution calculations separately. Supporting a causal claim would require additional interventions and controls; the associations described in the paper are not enough. These are evaluation criteria, not results already verified on new datasets.
The practical choice therefore depends on your objective, resources, and required level of certainty. **If the question is “what contribution patterns appear in these layers,” the SAE provides a specific instrument; if the question is “what causes an ability” or “what performance will it deliver in production,” the study does not answer that on its own.**
Checklist for your own evaluation
- 01Define whether you will measure reconstruction, contribution by expert, interpretive stability, or behavior after an intervention.
- 02Choose a layer and use the SAE trained for that same layer; do not interchange latent spaces.
- 03Describe the corpus, input preparation, and hardware so the results can be interpreted.
- 04Keep numerical metrics separate from automated explanations and test their stability with more than one sample.
- 05Do not turn correlations or expert labels into causal claims without additional interventions.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction