Mixture of experts

Evidence in the last complete seven days

Source mentions

6

Previous seven days 5

Distinct sources

3

Previous 2

Comparable growth

Not comparable

Ranking appearances

0

Collection observations and undated records are separate from dated publication counts. Observed in 7d：8 · Undated records：1

Daily history · up to 90 days

Bars show dated source mentions. The latest UTC day may be incomplete. Zero mentions and missing collection coverage are different conditions.

Dotted markers indicate days with no record. They are not counted as zero mentions.

Source evidence

Sources in the evidence shown：HF Daily Papers 2 · arXiv 6 · V2EX 2 · GitHub Trending 1 · GitHub Issues 3 · GitHub Releases 5 · Hugging Face 1 · Hacker News 1

1. HF Daily Papers · Published Nov 4, 2026

   UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling

   Universal image restoration aims to handle diverse degradations within a single model, but a reliable scaling strategy remains underexplored. Increasing backbone depth and width can lead to performance saturation, while degradation-aware specialization depends on accurate degradation estimation and is vulnerable to errors on out-of-distribution and real-worl

2. arXiv · Published Nov 4, 2026

   HybridSpec: Throughput-Aware Speculative Decoding for Hybrid GDN-MoE Models

   arXiv:2611.04055v1 Announce Type: new Abstract: In speculative decoding, parallel drafters such as DFlash make drafting cheap, so target verification dominates each round. A central choice is the verification length, the number of tokens that the target processes in each round. Verifying more draft tokens can accept more of them, but work after the first rej

   View source record →

3. V2EX · Published Nov 2, 2026

   [开源软件] 开源了一个 1600 行的大模型推理引擎，能跑真实模型，测试场景吞吐超过 vLLM

   想入门大模型推理，直接读 vLLM 太重。所以开源了 LiteMoE：约 1600 行 Python ，能跑真实的 Granite 模型，一个下午能读完主线。 适合想快速上手推理的同学： 交互演示，逐步看一次真实推理： https://zhulinsen.github.io/LiteMoE/visualizer/?lang=zh 教程，从 KV Cache 、MoE 讲到 DeepSeek 、Kimi 、Qwen 的新设计： https://github.com/ZhuLinsen/LiteMoE/blob/main/docs/tutorial.zh-CN.md 固定批次的测试场景下，吞吐是 vLLM 的 105 ～ 109%，脚本可复现。算子用的是 vLLM 的，提升来自算子外的执行层。定位是学习和实验，上线

   View source record →

4. GitHub Trending · Published Not supplied

   JustVugg/colibri

   A hierarchical inference engine for frontier MoE models: hot experts on the GPU (CUDA, Vulkan), warm ones in RAM, cold ones streamed from NVMe. Written in C, runs on the hardware you already own.

   View source record →

5. GitHub Issues · Published Nov 1, 2026

   [Doge] `DogeCDMoE` router `.view(2, bsz * seq_len, -1)` mixes tokens: with `is_moe=True`, logits depend on future tokens and on other sequences in the batch

   View source record →

6. arXiv · Published Oct 31, 2026

   LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

   arXiv:2610.37720v1 Announce Type: new Abstract: Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory

   View source record →

7. GitHub Releases · Published Oct 29, 2026

   v5.22.0: Mistral Large 4 (Le Chonk)

   New Models Mistral Large 4 (Le Chonk) Mistral Large 4 is Mistral's largest and most capable model to date: a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters i

   View source record →

8. HF Daily Papers · Published Oct 27, 2026

   Muon Makes the First Mixture-of-Experts Layer Harder to Balance

   Many open mixture-of-experts (MoE) language models train with Muon instead of AdamW. Most balance the load with a bias on each expert's routing score, adjusted after every step. Their recipes disagree on whether Muon should also train the router. In models with 55M and 108M active parameters, we measure how the optimizer changes the span of these biases, the

9. GitHub Releases · Published Oct 26, 2026

   Release: v5.21.0

   Release v5.21.0 New Model additions MiniCPM-OCR MiniCPM-OCR is a document-understanding model from OpenBMB's MiniCPM-V family. It keeps the MiniCPM-V 4.6 architecture, which pairs a SigLIP vision encoder and a window-attention merger with a Qwen3.5 language model. It is specialized for reading text-dense images such as scanned pages, screenshots, receipts an

   View source record →

10. arXiv · Published Oct 26, 2026

    Kolibri: A Sovereign European Model on the Pareto Frontier

    arXiv:2610.31945v1 Announce Type: new Abstract: This report introduces Kolibri, an English-German mixture-of-experts transformer with 78.1B total parameters and 3.46B active parameters per token, released as open weights under the Apache 2.0 licence. We built Kolibri for sovereign and specialised deployment, with a particular focus on German and on regulated

    View source record →

11. arXiv · Published Oct 26, 2026

    EDWARD-NMT: Elastic Dimensions With Adaptive Routing and Dual-consistency for Neural Machine Translation

    arXiv:2610.32115v1 Announce Type: new Abstract: Sparse Mixture-of-Experts (MoE) is a powerful paradigm for scaling Neural Machine Translation (NMT) and managing language heterogeneity. However, practical limitations hinder its deployment on resource-constrained devices: the entire expert pool must remain in memory, and the per-token compute budget is statica

    View source record →

12. GitHub Issues · Published Oct 22, 2026

    DiffusionGemma: generated text changes with batch size (MoE router projection and self-conditioning matmul)

    View source record →

13. arXiv · Published Oct 21, 2026

    Mixture of Phonetic Experts Based Low-Rank Adaptation of Conformer Models for Accented English Speech Recognition

    arXiv:2610.26888v1 Announce Type: new Abstract: We present Mixture of Phonetic Experts based Low-Rank Adaptation (MoPE-LoRA), a parameter-efficient adaptation framework for accented English speech recognition. Recent mixture-of-experts approaches typically allocate experts per accent, implicitly treating accent identity as the primary axis of variation. In c

    View source record →

14. GitHub Issues · Published Oct 19, 2026

    [Bug] Qwen3-MoE router auxiliary loss is multiplied by the DDP world size in Trainer

    View source record →

15. arXiv · Published Oct 19, 2026

    Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss

    View source record →

16. GitHub Releases · Published Oct 15, 2026

    Release v5.20.0

    Release v5.20.0 New Model additions Muse Spark Open Muse Spark Open is a multimodal mixture-of-experts model from Meta that takes text, images, audio, and video as input and generates text. It has 816B total parameters (42B active per token) and a context window of up to 1,048,576 tokens shared across modalities. It is trained for agentic coding tasks such a

    View source record →

17. Hugging Face · Published Oct 15, 2026

    Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

    View source record →

18. Hacker News · Published Oct 13, 2026

    Nvidia Nemotron 3.5 Bolt: open 74B MoE for local agents on RTX Spark

    View source record →

19. V2EX · Published Oct 13, 2026

    [推广] Moe - MCP 文字动画服务 送 Pro

    View source record →

20. GitHub Releases · Published Sep 10, 2026

    Release 5.17.0

    Release v5.17.0 New Model additions HYV4 Hy4-Preview is a 780B-parameter mixture-of-experts language model that activates 49B parameters per token. Each MoE layer holds 256 routed experts plus one always-active shared expert and routes every token to 8 of them. The context window is 1M tokens. The architecture combines four features: Multi-head Latent Attent

    View source record →

21. GitHub Releases · Published Aug 26, 2026

    Release: v5.16.0

    Release v5.16.0 New Model additions Qwen4-Exp Qwen4-Exp builds on Qwen3.5's hybrid text and multimodal architecture with three key components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE). GR is a Qwen-developed residual architecture that combines Hyper-Connection with GatedNorm. It mixes multiple residual streams with fine-

    View source record →

Review priority 74 / 100

This score orders research candidates. It is not search volume, a forecast or verified commercial demand.

• Evidence

  25.4 / 30

  Distinct dated articles or first observations; repeat crawling adds nothing.

• Publisher diversity

  24 / 25

  Conservative owner grouping; syndicated copies count once.

• Comparable change

  0 / 20

  Comparable complete weeks only; no percentage from a zero baseline.

• Recent evidence

  15 / 15

  Latest publication or first observation, not repeated fetch time.

• Name specificity

  9.7 / 10

  Explicit entities and technical concepts outrank ambiguous acronyms.

Limitations and interpretation

• First observation is a collection milestone, not proof that a word or product is new to the world.
• Counts describe monitored public sources, not the whole web or search demand.
• First observed here is not the invention date of a name or idea.
• Candidates come from public titles, summaries and entity URLs, not verified full-text analysis.
• Owner grouping is conservative; undisclosed syndication may remain.
• Undated observations cannot establish publication growth.
• Repeat ranking appearances are observations, not new articles or independent endorsements.
• An established concept or common product name retained as a baseline, not a newly invented term.
• Both weeks need adequate, comparable collection coverage.
• Sources behind this term changed or lack a comparable baseline.
• Recent evidence includes collection over three days after publication.