MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models

Abstract: Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

• The paper introduces MoEMB, a method that improves universal multimodal embeddings (UME) by scaling the encoder's capacity through sparse Mixture-of-Experts (MoE) routing, while maintaining high-efficiency resource bounds.
• By using a pretrained sparse multimodal MoE backbone, MoEMB achieves strong performance despite the model being comparably lightweight.
• The MoEMB system shows a ten-point accuracy gain on a clean benchmark, even when halving token transitions and activating fewer than 50% of the experts.

Universal multimodal embedding (UME) requires one encoder to support heterogeneous inputs and objectives, including image, video, visual-document, retrieval, classification, question answering, grounding, and reasoning-intensive search. “MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models” (2609.08663) addresses a specific capacity problem in this setting: increasing embedding dimensionality, storing multiple vectors, or generating intermediate reasoning tokens does not directly expand the encoder’s representational capacity. The paper instead scales the encoder through sparse mixture-of-experts (MoE) capacity while maintaining a bounded active-parameter and FLOP budget.

Problem formulation and central thesis

The paper treats a UME model as an instruction-conditioned multimodal encoder that maps a query or target to a normalized single-vector representation. Training uses a one-directional InfoNCE objective over paired examples, with in-batch negatives. This formulation makes two constraints especially important. First, contrastive quality depends strongly on large batches because the negative pool is formed within the batch. Second, every query and target must be encoded, so increasing dense model size directly increases both activation memory and inference cost.

The authors motivate MoE scaling through an empirical distinction between task interference and insufficient capacity. A jointly trained generalist underperforms task-specialist models, but the generalist-specialist gap becomes smaller as backbone capacity increases. For the dense 2B, dense 9B, and sparse Qwen3.5-35B-A3B backbones, the reported mean gaps are 3.02, 1.92, and 1.13 points, respectively. The measured router-gradient conflict is also small: the mean pairwise conflict is 0.020 and the maximum is 0.112. These observations support the paper’s claim that capacity constraints, rather than destructive router-gradient interference, are the dominant source of multi-task degradation in the evaluated setting.

MoEMB distinguishes encoder-capacity scaling from representation scaling and test-time reasoning scaling. Dimension-based methods alter the size of the stored vector; multi-vector methods increase retrieval-side storage and interaction cost; think-then-embed systems increase computation through generated intermediate tokens. MoEMB instead increases the total number of expert parameters while keeping only a fixed number of experts active per token.

Figure 1: MoEMB scales the encoder through sparse expert capacity rather than embedding dimensionality, multiple stored vectors, or intermediate reasoning tokens.

Native sparse MoE adaptation

The principal model is built by adapting pretrained sparse multimodal MoE backbones. Each MoE layer contains a shared expert and a large pool of routed experts. A token is assigned to its top-$k$ routed experts, whose outputs are combined with the shared expert output. MoEMB-A3B uses 256 routed experts, top-8 routing, 3.1B active parameters, 40 decoder layers, and a 2,048-dimensional hidden state. MoEMB-A10B uses the same number of routed experts, 10.2B active parameters, 48 decoder layers, and a 3,072-dimensional hidden state.

The paper compares three routes to MoE UME models. Native MoE reuses experts and routers learned during language-model pretraining. Dense-to-MoE upcycling replicates or partitions dense feed-forward networks and learns a new router during contrastive adaptation. Modality-level MoE assigns tokens from different modalities to corresponding expert groups rather than dynamically routing each token. Native MoE is clearly superior in the reported ablation:

Design	Active parameters	Overall MMEB-V2
Dense Qwen3.5-2B	2.2B	63.6
Dense-to-MoE upcycling	approximately 3.0B	63.1
Modality-level MoE	2.2B	64.2
Native sparse MoE, MoEMB-A3B	3.1B	70.4

The 6.8-point improvement over the dense baseline is therefore not attributable merely to adding an MoE wrapper. It depends on inheriting a router and expert decomposition established during large-scale pretraining. Upcycling fails to produce comparable gains: coarse top-1 upcycling is essentially equivalent to the dense baseline, while fine-grained partitioning reduces performance below 60 overall unless active computation is increased. The authors reasonably attribute this result to the inability of the limited contrastive dataset to learn a high-quality router from scratch.

A dedicated <emb> token is used for pooling. It improves the overall score from 69.50 with last-token pooling to 70.34. In contrast, router modification is consistently unhelpful. Freezing the pretrained router performs essentially identically to training it, while load-balancing and $z$-losses reduce performance to 66.62. Task-ID and instruction-conditioned router biases also degrade the result. The implication is strong and somewhat counter to standard downstream adaptation practice: the language-pretrained routing distribution transfers sufficiently well to contrastive UME training that additional routing regularization is harmful rather than beneficial.

Benchmark performance and expert scaling

MoEMB is evaluated on MMEB-V2, which contains 78 image, video, and visual-document tasks, and MRMR, which targets reasoning-intensive multimodal retrieval. The models are trained only on public MMEB-family data for the primary comparison. This distinction is important because several competing systems use external corpora that overlap with evaluation domains.

On MMEB-V2, MoEMB-A3B obtains an overall score of 70.4 with 3.1B active parameters. MoEMB-A10B reaches 72.4 with 10.2B active parameters. Among models trained only on public MMEB-family data, these results exceed the reported scores of reasoning-based systems despite substantially smaller active compute.

Model	Active parameters	MMEB-V2 overall	MRMR average
UME-R1-2B	2.2B	61.2	40.0
Embed-RL-4B	13.2B	68.1	46.8
UME-R1-7B	8.3B	65.6	48.2
MoEMB-A3B	3.1B	70.4	51.6
MoEMB-A10B	10.2B	72.4	—

The MRMR result is especially consequential for the paper’s comparison with think-then-embed methods. MoEMB-A3B scores 51.6, exceeding PLUME by 11.6 points, UME-R1-7B by 3.4 points, and Embed-RL-4B by 4.8 points, while activating only 3.1B parameters. These comparisons do not establish that MoE scaling universally dominates reasoning-based scaling, because the systems differ in backbone, training data, and optimization. They do establish that explicit autoregressive reasoning is not necessary to achieve the strongest reported performance under the paper’s public-data evaluation protocol.

The most direct evidence for the claimed scaling axis comes from varying the total number of experts while preserving top-8 activation. Increasing the expert pool from 32 to 256 raises MoEMB-A3B’s score by 10.4 points, while QPS decreases only from 34.2 to 29.3. Since the number of active experts remains fixed, total parameter capacity increases without a comparable increase in per-token computation.

The contrast with test-time reasoning is substantial. For Embed-RL-2B, enabling reasoning raises accuracy by only 1.5 points while reducing throughput from 41 to 0.11 QPS, approximately a 370-fold slowdown. The result supports the paper’s central systems argument: for high-volume embedding workloads, expert capacity can improve quality without imposing the throughput collapse associated with autoregressive reasoning.

Adaptive computation across tokens, layers, and experts

Sparse MoE routing bounds the number of active experts, but it still applies a fixed execution policy to every token. MoEMB therefore evaluates six adaptive-computation actions:

• token pruning;
• layer skipping;
• early exit;
• expert skipping;
• adaptive top-$k$ routing;
• permanent expert pruning.

The study includes both training-based controllers and inference-only heuristics. This breadth is important because a method that saves logical FLOPs may fail to improve wall-clock latency if the implementation does not physically compact sequences or exploit sparse expert execution.

Figure 2: The evaluated adaptive-computation actions include general token- and layer-level controls and MoE-specific expert execution controls.

The strongest individual trade-offs come from token pruning and adaptive top-$k$. At a representative operating point, uniform top-4 routing reduces FLOPs by 11.6%, improves QPS by $1.19\times$, and reduces the overall score by only 0.5 points. Keeping 50% of visual tokens with GDN-$\beta$ pruning reduces FLOPs by 22.8%, increases QPS by $1.39\times$, and reduces accuracy by 1.1 points. Permanently pruning the expert pool from 256 to 128 produces a 1.16-fold throughput improvement with a 0.8-point accuracy decrease, although this primarily reduces memory and total capacity rather than dynamic per-token FLOPs.

Action	Operating point	FLOPs reduction	QPS multiplier	Accuracy change
Uniform top-$k$	top-4 from top-8	11.6%	$1.19\times$	−0.5
GDN-$\beta$ token pruning	keep 50% visual tokens	22.8%	$z$0	−1.1
Trained layer skipping	skip 20% layers	approximately 20%	approximately $z$1	−0.8
REAP expert pruning	128 of 256 experts	approximately 0% dynamic FLOPs	$z$2	−0.8
MoDES plus GDN-$z$3	30% expert skip, 50% token keep	28.9%	$z$4	−1.2

The inference-only methods are often competitive with, and sometimes better than, trained controllers. A trained adaptive top-$z$5 model settles at 2.94 physical experts per token and obtains 67.2, whereas MoDES reaches 69.8 at a comparable budget. This result suggests that the pretrained router’s logit ordering already contains a useful estimate of expert importance. Randomly selecting four experts instead of using the router’s top four causes a 6.6-point degradation at the same FLOP reduction, demonstrating that the gain is not simply a consequence of reducing the number of experts.

Training becomes more useful at extreme token-pruning ratios. Inference-only GDN-$z$6 pruning performs competitively at moderate budgets, but at aggressive regimes that remove 70–75% of visual tokens, the trained dropper maintains roughly a 1.0-point loss while the heuristic incurs a 4.3-point loss. Conversely, several learned adaptive mechanisms are unstable. Gradual pruning with straight-through estimation and AViT-style halting fails to converge reliably, early-exit heads converge near the final layer, and some halting formulations drive almost all dropping to the earliest layer. The paper attributes these failures to high-variance credit assignment through discrete and delayed execution decisions.

Figure 3: Token pruning and adaptive top-$z$7 provide the most favorable budget–accuracy curves, while trained controllers become more useful at aggressive pruning budgets.

Combining adaptive actions is not automatically additive. Joint training under a single global FLOP budget strongly favors token dropping and reaches only 64.7 overall, a 5.6-point decline. Per-action simultaneous, alternating, and sequential strategies also perform worse than the strongest individual actions. The reported best combined policy is GRPO-based gradual dropping plus expert skipping, which reduces FLOPs by 22.5%, provides a $z$8 QPS improvement, and loses 0.7 points. This result indicates that adaptive actions compete for the same representational slack; independently imposing several target budgets can over-constrain the encoder.

GDN write strength as a training-free salience signal

A central technical contribution is the use of Gated DeltaNet write strength $z$9 for token importance. In a GDN layer, $k$0 controls the extent to which the recurrent state is updated by the current token. MoEMB averages this signal across heads and selected layers, then uses it to rank visual tokens without materializing attention maps.

This choice addresses a practical deployment constraint. FlashAttention and related fused kernels avoid storing full attention matrices, making attention-based token ranking inconvenient or incompatible with efficient inference. GDN-$k$1 is available directly from the recurrent update gate and therefore supplies a low-overhead signal.

The paper argues that $k$2 and attention measure complementary properties. Attention is sparse and highly variable, whereas $k$3 is smoother and more stable across depth and time. The reported mean token-wise coefficient of variation is 0.226 for $k$4 versus 2.316 for attention. Across layers, the average cosine similarity of $k$5 maps is 0.989, with a minimum of 0.973, compared with 0.752 for <emb> attention. Across video frames, the corresponding similarity is 0.992 for $k$6 and 0.744 for attention.

Figure 4: GDN write strength is concentrated in early layers and supports depth-structured allocation of token and expert computation.

Figure 5: GDN-$k$7 provides a smooth, early salience pattern, whereas <emb> attention is sparse, scattered, and increasingly selective at deeper layers.

The interpretation offered by the paper is that $k$8 identifies tokens that write substantially into the recurrent memory, rather than tokens that receive high instantaneous attention. This distinction explains why it remains spatially and temporally persistent. In image and document examples, $k$9 forms contiguous regions around semantically informative content; in video, it remains consistent across frames. The evidence supports $k$0 as a useful pruning heuristic, but not as a direct proof that write strength is an intrinsic measure of semantic importance. Its value is established operationally through the accuracy–compute results.

Adaptive computation is also modality- and depth-dependent. Token pruning occurs primarily in early layers and is more aggressive for images and videos, where spatial or temporal redundancy is greater. Visual documents retain more tokens because layout and text density are important. Text tokens are preserved early to establish task context. Expert skipping is more heterogeneous in early layers, while later layers converge toward similar skip rates across modalities. Adaptive top-$k$1 allocates more experts to text, particularly in early layers.

The routing analysis provides a complementary architectural explanation. Early layers show substantial modality specialization: within-modality routing distributions have a mean Jensen–Shannon distance of 0.114, compared with 0.323 between modalities. At layer 4, 172 of 256 experts exhibit strong modality preference, while at layer 38 only 43 do so and the inter-modality distance falls to 0.099. Image and video routing are especially similar, with a reported distance of 0.117. Thus, MoEMB uses modality-differentiated expert computation early and a more shared expert population deeper in the network.

Figure 6: Expert routing is modality-specialized in early layers and converges across modalities with depth.

Systems implementation and realized speedups

The paper explicitly separates logical FLOP reduction from deployed acceleration. For token pruning, the implementation gathers surviving tokens into packed variable-length sequences and updates cumulative sequence lengths so that downstream attention and feed-forward operations process fewer rows. Spatial masking alone provides almost no benefit because tensor dimensions remain unchanged.

At 50% token retention, packed variable-length execution reaches a $k$2 speedup, whereas masking remains near $k$3. Across the evaluated token-pruning configurations, packed execution converts 63–82% of removed FLOPs into measured speedup. For MoE execution, the authors use tile-aware grouped-GEMM kernels through SonicMoE. These kernels convert 56–58% of removed expert FLOPs into wall-clock savings, compared with 41% for unmodified grouped GEMM.

The analysis identifies two hardware constraints. First, grouped GEMM processes expert assignments in tiles, so removing tokens does not reduce latency until enough rows disappear to eliminate a tile. Second, reducing routed expert computation eventually reaches a floor determined by attention, shared experts, dispatch, and other non-routed operations. At shorter sequence lengths, aggressive expert skipping can therefore produce little or no speedup; at longer sequences, the same skip rate may remove complete tiles and yield an actual gain.

The paper also reports that a naive sequential expert loop is 5.7–8.0 times slower than fused grouped GEMM. Its apparent sensitivity to sparsity consequently overstates the practical benefit of expert skipping. This systems result is important: adaptive-computation claims must specify the execution backend, physical compaction strategy, sequence length, batch size, and kernel tiling regime.

Limitations and open questions

The strongest comparisons are restricted to models trained on public MMEB-family data. MoEMB remains behind systems trained with external data, so its reported state of the art is conditional on the data-regimen grouping. The model also does not support audio, limiting the scope of the claimed universality.

Several empirical conclusions are likewise bounded by the study design. The native MoE advantage is demonstrated using Qwen3.5 multimodal backbones and a particular contrastive adaptation protocol; it remains open whether routers pretrained under other objectives transfer equally well to UME. The conclusion that capacity dominates router-gradient conflict is supported by low measured conflict on the evaluated task categories, but may not hold for broader task mixtures or substantially different supervision. Similarly, GDN-$k$4 is evaluated on architectures containing Gated DeltaNet blocks; its portability to attention-only multimodal encoders is not established.

The wall-clock results depend on custom kernels and hardware-specific tiling behavior. The reported QPS improvements therefore should not be interpreted as architecture-independent constants. The combined adaptive-computation experiments also show that a nominal 50% logical-compute target does not necessarily yield a 50% latency reduction, and that independently trained controllers can interfere. Open questions include whether a unified controller can learn hardware-aware allocation directly, whether adaptive policies remain stable under larger batches and longer contexts, and whether audio integration changes the modality-specific routing and pruning patterns.

Conclusion

MoEMB proposes expert capacity as a scaling axis for universal multimodal embeddings. Its main empirical result is that adapting a pretrained sparse multimodal MoE is substantially more effective than upcycling a dense model or imposing modality-level experts. MoEMB-A3B reaches 70.4 on MMEB-V2 and 51.6 on MRMR with 3.1B active parameters, while increasing the total expert pool produces a 10.4-point gain with only a small QPS reduction.

The adaptive-computation study shows that MoE UME models contain exploitable redundancy at the token, expert, and layer levels. Uniform top-$k$5 reduction and GDN-$k$6 token pruning provide strong inference-only trade-offs, while packed sequence execution and tile-aware grouped-GEMM kernels are necessary to convert theoretical savings into measured throughput. The resulting picture is that MoE scaling supplies capacity without proportional active compute, and adaptive execution can further reduce that compute, but both benefits depend critically on pretrained routing structure and deployment-aware implementation.