Search Results for "moe"
Sort By:
Showing 47 open source projects for "moe"
View related business solutions-
Try Google Cloud Risk-Free With $300 in CreditUse your credit across every product. Compute, storage, AI, analytics. When it runs out, 20+ products stay free. You only pay when you choose to.
-
Gemini 3 and 200+ AI Models on One PlatformBuild generative AI apps with Vertex AI. Switch between models without switching platforms.
-
1
moe
A command line based editor inspired by Vim. Written in Nim
Moe is a lightweight, modern text editor written in Nim, designed for simplicity and efficiency while providing a minimal yet functional editing environment. -
2
Flash-MoE
Running a big model on a small laptop
...Overall, flash-moe represents a technical advancement in making MoE models more practical and deployable. -
3
DeepEP
DeepEP: an efficient expert-parallel communication library
DeepEP is a communication library designed specifically to support Mixture-of-Experts (MoE) and expert parallelism (EP) deployments. Its core role is to implement high-throughput, low-latency all-to-all GPU communication kernels, which handle the dispatching of tokens to different experts (or shards) and then combining expert outputs back into the main data flow. Because MoE architectures require routing inputs to different experts, communication overhead can become a bottleneck — DeepEP addresses that by providing optimized GPU kernels and efficient dispatch/combining logic. ... -
4
DeepSeek MoE
Towards Ultimate Expert Specialization in Mixture-of-Experts Language
DeepSeek-MoE (“DeepSeek MoE”) is the DeepSeek open implementation of a Mixture-of-Experts (MoE) model architecture meant to increase parameter efficiency by activating only a subset of “expert” submodules per input. The repository introduces fine-grained expert segmentation and shared expert isolation to improve specialization while controlling compute cost. -
Custom VMs From 1 to 96 vCPUs With 99.95% UptimeLive migration and automatic failover keep workloads online through maintenance. One free e2-micro VM every month.
-
5
HunyuanImage-3.0
A Powerful Native Multimodal Model for Image Generation
...It unifies multimodal understanding and generation in a single autoregressive framework, combining text and image modalities seamlessly rather than relying on separate image-only diffusion components. It uses a Mixture-of-Experts (MoE) architecture with many expert subnetworks to scale efficiently, deploying only a subset of experts per token, which allows large parameter counts without linear inference cost explosion. The model is intended to be competitive with closed-source image generation systems, aiming for high fidelity, prompt adherence, fine detail, and even “world knowledge” reasoning (i.e. leveraging context, semantics, or common sense in generation). ... -
6
Ling
Ling is a MoE LLM provided and open-sourced by InclusionAI
Ling is a Mixture-of-Experts (MoE) large language model (LLM) provided and open-sourced by inclusionAI. The project offers different sizes (Ling-lite, Ling-plus) and emphasizes flexibility and efficiency: being able to scale, adapt expert activation, and perform across a range of natural language/reasoning tasks. Example scripts, inference pipelines, and documentation. -
7
Kimi K2
Kimi K2 is the large language model series developed by Moonshot AI
Kimi K2 is Moonshot AI’s advanced open-source large language model built on a scalable Mixture-of-Experts (MoE) architecture that combines a trillion total parameters with a subset of ~32 billion active parameters to deliver powerful and efficient performance on diverse tasks. It was trained on an enormous corpus of over 15.5 trillion tokens to push frontier capabilities in coding, reasoning, and general agentic tasks while addressing training stability through novel optimizer and architecture design strategies. ... -
8
Ling-V2
Ling-V2 is a MoE LLM provided and open-sourced by InclusionAI
Ling-V2 is an open-source family of Mixture-of-Experts (MoE) large language models developed by the InclusionAI research organization with the goal of combining state-of-the-art performance, efficiency, and openness for next-generation AI applications. It introduces highly sparse architectures where only a fraction of the model’s parameters are activated per input token, enabling models like Ling-mini-2.0 to achieve reasoning and instruction-following capabilities on par with much larger dense models while remaining significantly more computationally efficient. ... -
9
Ring
Ring is a reasoning MoE LLM provided and open-sourced by InclusionAI
Ring is a reasoning Mixture-of-Experts (MoE) large language model (LLM) developed by inclusionAI. It is built from or derived from Ling. Its design emphasizes reasoning, efficiency, and modular expert activation. In its “flash” variant (Ring-flash-2.0), it optimizes inference by activating only a subset of experts. It applies reinforcement learning/reasoning optimization techniques. -
MongoDB Atlas runs apps anywhereMongoDB Atlas gives you the freedom to build and run modern applications anywhere—across AWS, Azure, and Google Cloud. With global availability in over 115 regions, Atlas lets you deploy close to your users, meet compliance needs, and scale with confidence across any geography.
-
10
Wan2.2
Wan2.2: Open and Advanced Large-Scale Video Generative Model
Wan2.2 is a major upgrade to the Wan series of open and advanced large-scale video generative models, incorporating cutting-edge innovations to boost video generation quality and efficiency. It introduces a Mixture-of-Experts (MoE) architecture that splits the denoising process across specialized expert models, increasing total model capacity without raising computational costs. Wan2.2 integrates meticulously curated cinematic aesthetic data, enabling precise control over lighting, composition, color tone, and more, for high-quality, customizable video styles. ... -
11
DeepSeek-V3
Powerful AI language model (MoE) optimized for efficiency/performance
DeepSeek-V3 is a robust Mixture-of-Experts (MoE) language model developed by DeepSeek, featuring a total of 671 billion parameters, with 37 billion activated per token. It employs Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture to enhance computational efficiency. The model introduces an auxiliary-loss-free load balancing strategy and a multi-token prediction training objective to boost performance. -
12
LLaMA-MoE
Building Mixture-of-Experts from LLaMA with Continual Pre-training
LLaMA-MoE is an open-source project that builds mixture-of-experts language models from LLaMA through expert partitioning and continual pre-training. The repository is centered on making MoE research more accessible by offering smaller and more affordable models with only about 3.0 to 3.5 billion activated parameters, which helps reduce deployment and experimentation costs. -
13
DeepSeek R1
Open-source, high-performance AI model with advanced reasoning
DeepSeek-R1 is an open-source large language model developed by DeepSeek, designed to excel in complex reasoning tasks across domains such as mathematics, coding, and language. DeepSeek R1 offers unrestricted access for both commercial and academic use. The model employs a Mixture of Experts (MoE) architecture, comprising 671 billion total parameters with 37 billion active parameters per token, and supports a context length of up to 128,000 tokens. DeepSeek-R1's training regimen uniquely integrates large-scale reinforcement learning (RL) without relying on supervised fine-tuning, enabling the model to develop advanced reasoning capabilities. ... -
14
Step 3.5 Flash
Fast, Sharp & Reliable Agentic Intelligence
Step 3.5 Flash is a cutting-edge, open-source large language model developed by StepFun-AI that pushes the frontier of efficient reasoning and “agentic” intelligence in a way that makes powerful AI accessible beyond proprietary black boxes. Unlike dense models that activate all their parameters for every token, Step 3.5 Flash uses a sparse Mixture-of-Experts (MoE) architecture that selectively engages only about 11 billion of its roughly 196 billion total parameters per token, delivering high-quality reasoning and interaction at far lower compute cost and latency than traditional large models. Its design targets deep reasoning, long-context handling, coding, and real-time responsiveness, making it suitable for building autonomous agents, advanced assistants, and long-chain cognitive workflows without sacrificing performance. -
15
EPLB
Expert Parallelism Load Balancer
...The logic is implemented in eplb.py and supports predicting placements given estimated expert usage weights. EPLB aims to reduce hot-spotting and ensure more uniform usage of compute resources in large MoE deployments. -
16
DeepSeekMath-V2
Towards self-verifiable mathematical reasoning
...Unlike general-purpose LLMs that might generate plausible-looking math but sometimes hallucinate or mishandle rigorous logic, Math-V2 is engineered to not only generate solutions but also self-verify them, meaning it examines the derivations, checks logical consistency, and flags or corrects mistakes, producing proofs + verification rather than just a final answer. Under the hood, Math-V2 uses a massive Mixture-of-Experts (MoE) architecture (activated parameter count reportedly in the hundreds of billions) derived from DeepSeek’s experimental base architecture. For math problems, it employs a generator-verifier loop: it first generates a candidate proof (or solution path), then runs a verifier that assesses correctness and completeness. -
17
MiniMax-M2
MiniMax-M2, a model built for Max coding & agentic workflows
MiniMax-M2 is an open-weight large language model designed specifically for high-end coding and agentic workflows while staying compact and efficient. It uses a Mixture-of-Experts (MoE) architecture with 230 billion total parameters but only 10 billion activated per token, giving it the behavior of a very large model at a fraction of the runtime cost. The model is tuned for end-to-end developer flows such as multi-file edits, compile–run–fix loops, and test-validated repairs across real repositories and diverse programming languages. ... -
18
Profile Data
Analyze computation-communication overlap in V3/R1
profile-data is a repository that publishes profiling traces and metrics from DeepSeek’s training and inference infrastructure (especially during DeepSeek-V3 / R1 experiments). The profiling data targets insights into computation-communication overlap, pipeline scheduling (e.g. DualPipe), and how MoE / EP / parallelism strategies interact in real systems. The repository contains JSON trace files like train.json, prefill.json, decode.json, and associated assets. Users can load them into tools like Chrome tracing to inspect GPU idle times, overlapping operations, and scheduling alignment. The idea is to bring transparency to internal efficiency tradeoffs, enabling researchers to reproduce, analyze, or improve on DeepSeek’s parallelism strategies. ... -
19
Qwen3-Omni
Qwen3-omni is a natively end-to-end, omni-modal LLM
Qwen3-Omni is a natively end-to-end multilingual omni-modal foundation model that processes text, images, audio, and video and delivers real-time streaming responses in text and natural speech. It uses a Thinker-Talker architecture with a Mixture-of-Experts (MoE) design, early text-first pretraining, and mixed multimodal training to support strong performance across all modalities without sacrificing text or image quality. The model supports 119 text languages, 19 speech input languages, and 10 speech output languages. It achieves state-of-the-art results: across 36 audio and audio-visual benchmarks, it hits open-source SOTA on 32 and overall SOTA on 22, outperforming or matching strong closed-source models such as Gemini-2.5 Pro and GPT-4o. ... -
20
DeepGEMM
Clean and efficient FP8 GEMM kernels with fine-grained scaling
...The library is designed to work cleanly and simply, avoiding overly templated or heavily abstracted code, while still delivering performance that rivals expert-tuned libraries. It supports both standard and “grouped” GEMMs, which is useful for architectures like Mixture of Experts (MoE) that require segmented matrix multiplications. One distinguishing aspect is that DeepGEMM compiles its kernels at runtime (via a lightweight Just-In-Time (JIT) module), so users don’t need to precompile CUDA kernels before installation. Despite its lean design, it includes scaling strategies (fine-grained scaling) and optimizations inspired by cutting edge systems (drawing from ideas in CUTLASS, CuTe) but in a more streamlined form. -
21
Qwen3
Qwen3 is the large language model series developed by Qwen team
Qwen3 is a cutting-edge large language model (LLM) series developed by the Qwen team at Alibaba Cloud. The latest updated version, Qwen3-235B-A22B-Instruct-2507, features significant improvements in instruction-following, reasoning, knowledge coverage, and long-context understanding up to 256K tokens. It delivers higher quality and more helpful text generation across multiple languages and domains, including mathematics, coding, science, and tool usage. Various quantized versions,... -
22
Wan2.1
Wan2.1: Open and Advanced Large-Scale Video Generative Model
Wan2.1 is a foundational open-source large-scale video generative model developed by the Wan team, providing high-quality video generation from text and images. It employs advanced diffusion-based architectures to produce coherent, temporally consistent videos with realistic motion and visual fidelity. Wan2.1 focuses on efficient video synthesis while maintaining rich semantic and aesthetic detail, enabling applications in content creation, entertainment, and research. The model supports... -
23
MiniMax-01
Large-language-model & vision-language-model based on Linear Attention
MiniMax-01 is the official repository for two flagship models: MiniMax-Text-01, a long-context language model, and MiniMax-VL-01, a vision-language model built on top of it. MiniMax-Text-01 uses a hybrid attention architecture that blends Lightning Attention, standard softmax attention, and Mixture-of-Experts (MoE) routing to achieve both high throughput and long-context reasoning. It has 456 billion total parameters with 45.9 billion activated per token and is trained with advanced parallel strategies such as LASP+, varlen ring attention, and Expert Tensor Parallelism, enabling a training context of 1 million tokens and up to 4 million tokens at inference. ... -
24
MiMo-V2-Flash
MiMo-V2-Flash: Efficient Reasoning, Coding, and Agentic Foundation
MiMo-V2-Flash is a large Mixture-of-Experts language model designed to deliver strong reasoning, coding, and agentic-task performance while keeping inference fast and cost-efficient. It uses an MoE setup where a very large total parameter count is available, but only a smaller subset is activated per token, which helps balance capability with runtime efficiency. The project positions the model for workflows that require tool use, multi-step planning, and higher throughput, rather than only single-turn chat. Architecturally, it highlights attention and prediction choices aimed at accelerating generation while preserving instruction-following quality in complex prompts. ... -
25
Kimi K2.5
Moonshot's most powerful AI model
Kimi K2.5 is Moonshot AI’s open-source, native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed vision and text tokens. Based on a 1T-parameter Mixture-of-Experts (MoE) architecture with 32B activated parameters, it integrates advanced language reasoning with strong visual understanding. K2.5 supports both “Thinking” and “Instant” modes, enabling either deep step-by-step reasoning or low-latency responses depending on the task. Designed for agentic workflows, it features an Agent Swarm mechanism that decomposes complex problems into coordinated sub-agents executing in parallel. ...
