MoM-open: Cross-Domain Disambiguation

• MoM-open is an ambiguous term that spans queueing theory, machine learning, and vision, with meanings varying by context.
• In queueing theory, the Method of Moments is applied exclusively to closed systems, with no open-network extension available.
• In ML and vision, MoM-open refers to open-source implementations like Mixture-of-Experts models and monocular motion estimation techniques.

The available literature suggests that “MoM-open” is not a standardized term, but rather an ambiguous label whose meaning depends strongly on field and context. In queueing theory, the most direct “MoM” reference is the Method of Moments for closed multiclass product-form queueing networks, and the relevant paper is explicit that it does not provide an extension to open networks (0902.3065). In contemporary machine learning systems, similar strings are used for open Mixture-of-Experts LLMs, Mixture-of-Models routing, Mixture-of-Memories sequence models, scenario-aware document memories for RAG, and open-source motion-estimation implementations (Xue et al., 2024). As a result, the phrase is best understood as a cross-disciplinary shorthand rather than a single established technical object.

1. Term status and principal interpretations

In the cited arXiv literature, “MoM” appears in several non-equivalent senses. One usage is the queueing-theoretic Method of Moments, where the key development is the Multi-Branched Method of Moments for exact analysis of closed multiclass product-form queueing networks (0902.3065). Another usage belongs to robust statistics and learning, where MOM denotes Median-of-Means, as in robust classification and robust Wasserstein estimation (Lecué et al., 2018). A third usage appears in lattice QCD, where RI/MOM and RI/SMOM denote momentum-subtraction renormalization schemes rather than an “open” systems paradigm (Bi et al., 2017).

A different cluster of meanings is tied to open software, open-weight models, or open deployment. OpenMoE is described as a family of fully open-sourced and reproducible decoder-only MoE LLMs (Xue et al., 2024). Brick operates in the Mixture-of-Models (MoM) paradigm and is explicitly motivated by routing between heterogeneous model pools, including local or open-weight models and stronger external models (Massa et al., 11 Jun 2026). MoM: Linear Sequence Modeling with Mixture-of-Memories and MoM: Mixtures of Scenario-Aware Document Memories both have open code releases, but they address very different technical problems: one is a linear sequence model with multiple memory states, the other a document-memory framework for retrieval-augmented generation (Du et al., 19 Feb 2025). This plurality of meanings is the main reason the phrase “MoM-open” requires explicit disambiguation.

2. Queueing-theoretic MoM and the absence of an open-network extension

In queueing theory, the most relevant source is “The Multi-Branched Method of Moments for Queueing Networks” (0902.3065). Its problem setting is unambiguous: it studies closed queueing networks only, with multiple classes, under product-form assumptions, using normalizing constants of the underlying Markov chain. The paper introduces a stronger recursion for the Method of Moments by incorporating the generalized convolution expression, which yields a multi-branched recursion over models with different numbers of queues (0902.3065).

The central technical distinction from Mean Value Analysis is that MoM recursively computes higher-order moments / normalizing constants rather than only mean values. The paper states the core exact relations as the convolution expression

$$G(\vec m+\vec 1_k,\vec N)=G(\vec m,\vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$

the population constraint

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$

and the generalized convolution expression

$$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$

By adding GCEs, the method shrinks the working basis and achieves large computational savings; the paper reports several cases where the proposed algorithm is between 1,000 and 10,000 times faster and less memory consuming than the original MoM (0902.3065).

For the specific reading “MoM-open” as “MoM for open product-form queueing networks,” the paper’s answer is explicit: no. The formulation depends on fixed class populations $\vec N$, on normalizing constants $G(\vec m,\vec N)$, and on population constraints specific to the closed regime. The paper states that there is no extension to open or mixed networks in the paper (0902.3065). This makes the queueing interpretation of “MoM-open” chiefly a negative one: the relevant MoM paper is foundational for closed multiclass exact analysis, but it does not define an open-network counterpart.

3. Open model systems: OpenMoE and Mixture-of-Models routing

In LLM systems, the “open” reading is more literal. OpenMoE is presented as an early effort on open Mixture-of-Experts LLMs, with code, data pipeline, and intermediate checkpoints released for the open-source community (Xue et al., 2024). The paper describes fully open-sourced and reproducible decoder-only MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T tokens, and it emphasizes three routing findings: Context-Independent Specialization, Early Routing Learning, and Drop-towards-the-End (Xue et al., 2024). Architecturally, OpenMoE follows an ST-MoE-style design with residual-MoE blocks and top-2 token-choice routing,

$$\mathrm{MoE}(x)=\sum_{i=1}^E g(x)_i\, e_i(x), \qquad g(x)=\mathrm{TopK}(\mathrm{softmax}(f(x))),$$

with $K=2$ (Xue et al., 2024).

A related but distinct “MoM-open” usage appears in Mixture-of-Models routing. Brick is a router for the MoM paradigm, where the routed unit is an entire pretrained model rather than an internal expert (Massa et al., 11 Jun 2026). Brick scores each model on six capability dimensions, combines this with a per-query difficulty estimate, and dispatches via a cost-penalized geometric rule (Massa et al., 11 Jun 2026). Its routing objective is based on asymmetric residuals in capability space and a cost term,

$$D_m = \sqrt{\sum_c\!\left(u_{m,c}^2 + \lambda\,o_{m,c}^2\right)}, \qquad J_m = D_m + \beta\,a_m, \qquad m^{\star}=\arg\min_m J_m.$$

On a benchmark of 5,504 queries, Brick at max-quality reaches 76.98\% accuracy, while at a neutral cost-quality profile it reaches 74.11\% accuracy at 4.71x lower cost than always using the strongest model (Massa et al., 11 Jun 2026).

These two lines of work show that, in current ML systems, “MoM-open” often points to open-weight or open-deployment infrastructure rather than to the queueing-theoretic Method of Moments. OpenMoE focuses on transparent sparse-model training and routing diagnostics, whereas Brick treats MoM as a deployment-time routing problem over heterogeneous model pools (Xue et al., 2024).

4. Memory-centric MoM architectures

Another major contemporary meaning is MoM as Mixture-of-Memories. “MoM: Linear Sequence Modeling with Mixture-of-Memories” introduces a linear sequence architecture that replaces the usual single fixed-size memory state with multiple independent memory states, with a router network directing input tokens to specific memory states (Du et al., 19 Feb 2025). The paper begins from the recurrent form of linear attention,

$$\bm M_t = \bm M_{t-1} + \bm k_t^T\bm v_t, \qquad \bm o_t = \bm q_t \bm M_t,$$

and generalizes it to a routed family $\{\bm M_t^1,\dots,\bm M_t^M\}$ with top-$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$0 selection,

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$1

The paper reports that MoM significantly outperforms current linear sequence models on downstream language tasks, particularly recall-intensive tasks, while preserving linear-complexity training and constant-complexity inference per token; its code is released at OpenSparseLLMs/MoM and OpenSparseLLMs/Linear-MoE (Du et al., 19 Feb 2025).

A different memory-oriented use appears in “MoM: Mixtures of Scenario-Aware Document Memories for Retrieval-Augmented Generation Systems” (Zhao et al., 16 Oct 2025). Here MoM is a document-processing framework that transforms RAG preprocessing from passive chunking to proactive understanding. A document is converted into

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$2

where $$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$3 is the outline, $$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$4 the core content, and $$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$5 the atomic chunks (Zhao et al., 16 Oct 2025). Candidate memories are generated by multi-path sampling and ranked by Atomic Chunks Clarity

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$6

and Core Content Completeness

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$7

then fused by reciprocal rank fusion. The paper explicitly provides an open-source code release at https://github.com/MemTensor/MoM, but it also states that it does not explicitly mention a term or variant called “MoM-open” (Zhao et al., 16 Oct 2025).

These memory-centric MoM papers share the theme of structured memory construction, but they solve very different problems: one concerns recurrent sequence modeling, the other document-memory extraction for RAG. Their relevance to “MoM-open” lies primarily in open implementation and in the reuse of the acronym MoM (Du et al., 19 Feb 2025).

5. Motion-related open implementations

A separate interpretation comes from motion estimation. “Momo: Monocular Motion Estimation on Manifolds” is likely the most direct source when “MoM-open” is used to refer to an open implementation of a named algorithm (Graeter et al., 2017). The paper explicitly states that the authors publish the code on GitHub and describes Momo as a monocular frame-to-frame motion estimation methodology for visual odometry that uses a vehicle motion model to constrain estimation on a lower-dimensional manifold (Graeter et al., 2017). Its preferred residual is the AnglePlane objective,

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$8

and the multi-camera formulation jointly optimizes

$$N_r G(\vec m, \vec N)=Z_{r} G(\vec m, \vec N-\vec 1_r)+\sum_{k=1}^{M} m_kD_{k,r} G(\vec m+\vec 1_k, \vec N-\vec 1_r),$$9

The paper reports real-time runtime of roughly 5–20 ms, multi-camera support even without overlap, and robust operation with only 100–300 feature matches (Graeter et al., 2017).

A broader open-motion-planning reading is represented by “The Open Motion Planning Library 2.0”, although that paper does not use MoM as its core acronym (Guo et al., 28 May 2026). It describes OMPL 2.0 as an open-source infrastructure for sampling-based planning, with state spaces, validators, samplers, and planners as core abstractions, and with support for asymptotically optimal planning, lazy planning, constrained planning, and temporal-logic goals (Guo et al., 28 May 2026). This is relevant only under a looser “open motion” interpretation of the query, not as a canonical definition of “MoM-open.”

6. Other MOM traditions and the need for disambiguation

Beyond queueing, model routing, memory architectures, and motion estimation, the acronym MOM has well-established meanings that are unrelated to any “open” variant. In robust statistics and learning, Median-of-Means appears in robust Wasserstein estimation and robust classification. “When OT meets MoM: Robust estimation of Wasserstein Distance” introduces MoM- and MoU-based estimators for contaminated optimal transport, including

$$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$0

and connects them to robust WGAN training (Staerman et al., 2020). “Robust classification via MOM minimization” defines MOM empirical risk

$$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$1

and studies classification under arbitrary outliers (Lecué et al., 2018). In these works, “open” is not part of the method name.

In lattice QCD, RI/MOM and RI/SMOM denote momentum-subtraction renormalization schemes rather than open systems. The overlap-bilinear studies compare exceptional-momentum RI/MOM,

$$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$2

with symmetric non-exceptional RI/SMOM,

$$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$3

and show that both schemes can yield consistent $$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$4 renormalization constants after careful $$G(\vec m,\vec N)=G(\vec m-\vec 1_k, \vec N)+\sum_{r=1}^{R} D_{k,r} G(\vec m, \vec N-\vec 1_r).$$5 extrapolation (He et al., 2021). A related quasi-PDF paper argues that RI/MOM for Wilson-line operators leaves a residual linear divergence in lattice regularization (Zhang et al., 2020). These usages are technically important but should not be conflated with “open” model or software interpretations.

The most reliable way to read “MoM-open,” therefore, is by domain-specific expansion. In queueing theory it most naturally asks whether the Method of Moments has an open-network extension, to which the relevant answer is negative (0902.3065). In modern ML systems it often points to open-weight MoE models, Mixture-of-Models routing, or open implementations of MoM architectures (Xue et al., 2024). In computer vision and robotics it may refer to the open-source implementation of Momo (Graeter et al., 2017). The phrase is thus best treated as a disambiguation problem rather than as the name of a single unified method.