实体

模型、机构、产品与重要对象,共 34 个页面。列表展示摘要、主要小节和标签,方便快速判断是否值得进入详情。

Composer 2

2026-04-15

Cursor Research Team 发布的专用 agentic software engineering 模型(2026-03-24)。

基础架构训练流程阶段一:Continued Pretraining阶段二:Reinforcement Learning
model architecture training inference company open-source

DDTree

2026-04-21

DDTree 是 DFlash 的扩展方法,通过构建 Draft Tree 从 block diffusion 的 per-position distributions 中选择多条候选路径,进一步提升 acceptance length。

背景:DFlash 的局限DDTree 的创新Draft Tree from Block Diffusion DistributionsBest-First Heap Algorithm
inference-optimization speculative-decoding diffusion tree-based

DFlash

2026-04-21

DFlash 是一种基于 Block Diffusion 的 Speculative Decoding 框架,用轻量级 diffusion model 替代传统 autoregressive drafting,实现并行 token 生成,达到 6x+ 无损加速。

核心问题DFlash 的创新核心思想:The Target Knows BestBlock Diffusion Drafting
inference-optimization speculative-decoding diffusion acceleration

Doubao App

2026-04-15

豆包 (Doubao) is ByteDance's AI assistant app, serving hundreds of millions of users in China. It is the deployment platform for the seeduplex full-duplex speech model, where the "打电话" (phone call) feature provides native full-duplex voice interaction. Doubao represents one of the largest-scale deployments of end-to-end ...

Key FactsRelatedReferences
voice-assistant chatbot company

EMOVA

2026-04-15

CVPR 2025,HKUST/SenseTime 联合发布的端到端 Omni-modal LLM,首个同时在视觉-语言和语音基准上达到 SOTA 的模型。

核心创新关键结果局限关联
model multimodal speech-model architecture open-source codec

Freeze-Omni

2026-04-09

腾讯团队于 2024 年 11 月发布的语音对话模型,核心创新是保持 LLM 参数冻结的同时实现 speech-to-speech 对话和 duplex 能力。

基本信息架构核心优势相关页面
model speech-model end-to-end company

FunASR — Fundamental End-to-End Speech Recognition Toolkit

2026-05-01

FunASR is an open-source end-to-end speech recognition toolkit developed by the Speech Lab of DAMO Academy, Alibaba Group. Its mission is to bridge the gap between academic research and industrial applications by providing pre-trained models on large-scale industrial corpora, comprehensive training/finetuning pipelines...

OverviewArchitecture & DesignCore Pipeline Stages (run.sh)Key Modules
model speech-model open-source end-to-end streaming

FunAudioLLM — Voice Understanding & Generation Foundation Models

2026-05-01

FunAudioLLM is a model family designed to enable natural voice interactions between humans and large language models. Developed by Alibaba Group's Tongyi SpeechTeam (formerly DAMO Speech Lab), it unifies two foundation models — sensevoice for voice understanding and cosyvoice for voice generation — built on top of the ...

OverviewFamily ArchitecturePipeline IntegrationCore Models
model speech-model open-source end-to-end

Kimi K2

2026-04-15

Kimi K2 is a large language model developed by Moonshot AI (月之暗面). It serves as the base model for composer2, the Cursor Composer 2 agentic coding model, which is built on a Kimi K2.5 1.04T/32B MoE architecture. Kimi K2 represents Moonshot AI's entry into the competitive LLM landscape.

Key FactsRelatedReferences
model open-source

LLaMA-Omni

2026-04-09

中科院团队基于 LLaMA-3.1-8B 构建的语音交互模型,发表于 ICLR 2025。

基本信息架构贡献相关页面
model speech-model end-to-end open-source

LongCat-Flash

2026-05-01

美团 LongCat 团队发布的旗舰级开源 MoE 大语言模型(arXiv 2509.01322,2025-09)。560B 总参数,18.6B~31.3B 动态激活(平均 27B),专为计算效率和 Agentic 能力设计。

一、核心参数二、架构创新Zero-Computation Experts(零计算专家)Shortcut-Connected MoE (ScMoE)
model architecture training inference alignment open-source

LongCat-Flash-Omni

2026-04-27

美团 LongCat 团队发布的全模态开源模型(arXiv 2511.00279,2025-11)。560B 总参数,27B 激活参数,在开源模型中达到 SOTA 级别的全模态能力,同时保持强单模态性能。

核心参数架构设计整体架构:端到端统一框架ScMoE (Shortcut-connected MoE)
model architecture multimodal speech-model training inference

LongCat-Flash-Thinking

2026-05-01

美团 LongCat 团队发布的开源 MoE 推理模型(arXiv 2509.18883,2025-09)。基于 longcat-flash 560B 架构,通过长 CoT 冷启动 + 大规模 RL 两阶段训练,在复杂推理和 Agentic 任务上达到开源 SOTA。

一、核心参数二、训练管线(两阶段)阶段一:Long CoT 冷启动训练阶段二:大规模强化学习(RL)
model architecture training reasoning chain-of-thought tool-use

LongCat-Flash-Thinking-2601

2026-05-01

美团 LongCat 团队发布的升级版开源 MoE 推理模型(arXiv 2601.16725,2026-01)。在 longcat-flash-thinking 基础上,通过环境扩展、噪声感知训练和 Heavy Thinking 模式,大幅提升 Agentic 推理能力。

一、核心参数二、训练框架预训练中期训练(Mid-training)
model architecture training reasoning tool-use open-source

LongCat-Image

2026-05-01

美团 LongCat 团队发布的开源双语图像生成基础模型(arXiv 2512.07584,2025-12)。6B 参数,专攻中英双语文字渲染、照片级真实感和高效部署,性能优于 20B+ 的 MoE 架构模型。

一、核心参数二、架构设计混合架构设计理念
model architecture training open-source multimodal

LongCat-Next

2026-05-01

美团 LongCat 团队发布的原生多模态自回归模型(arXiv 2603.27538,2026-03)。核心创新是 DiNA(Discrete Native Autoregression)框架——将所有模态统一离散化为 token,在单一共享离散空间中实现真正的原生多模态建模。

一、核心理念问题:语言中心主义的多模态方案:DiNA 框架二、架构创新
model architecture training open-source multimodal reasoning

LongCat-Video

2026-05-01

美团 LongCat 团队发布的开源视频生成基础模型(arXiv 2510.22200,2025-10)。13.6B 参数,基于 DiT 框架,支持文生视频、图生视频、视频续写,通过多奖励 RLHF 达到 SOTA 水平。

一、核心参数二、架构设计统一架构Block Sparse Attention
model architecture training open-source multimodal

Mem0: AI Agent 的可扩展长期记忆层

2026-04-17

Mem0(读作 "mem-zero")是一个为 AI Agent 提供智能长期记忆层的开源框架,

概述核心架构Mem0 基础版:Incremental Memory PipelineMem0^g:Graph-based Memory
company open-source memory agent long-term graph

MGM-Omni

2026-04-15

CUHK/JiaLiab 发布的 "Brain-Mouth" 双轨 Omni LLM,专注长音频理解和个性化语音生成。

核心架构关键创新性能关联
model multimodal speech-model architecture streaming codec

MinMo

2026-04-09

阿里团队于 2025 年 1 月发布的无缝语音交互多模态大模型。

基本信息贡献相关页面
model speech-model multimodal company

Moshi

2026-04-09

Kyutai 实验室于 2024 年 9 月发布的首个实时全双工语音对话模型。开创了多流自回归架构范式。

基本信息架构意义相关页面
model speech-model full-duplex open-source end-to-end

Qwen3

2026-04-10

Qwen3 is the latest series of open-weight large language models (LLMs) from the Qwen model family, released under Apache 2.0. It includes both dense and Mixture-of-Experts (MoE) architectures with parameter scales ranging from 0.6B to 235B.

Key InnovationsModel VariantsDense ModelsMoE Models
model architecture open-source

Qwen3-ASR

2026-04-15

阿里 Qwen 团队 2026-01 发布的多语言 ASR 模型家族,基于 qwen3-omni 的音频理解能力构建。

模型家族训练流程 (四阶段)语言支持性能亮点
model speech-model architecture open-source benchmark inference

Qwen3-Omni

2026-04-15

阿里 Qwen 团队 2025-09 发布的统一多模态模型,首次实现 text/image/audio/video 全模态 SOTA 且无单模态退化。

核心架构: Thinker-Talker MoE关键设计决策性能指标三阶段预训练
model multimodal speech-model architecture open-source streaming

Qwen3-TTS

2026-04-15

阿里 Qwen 团队 2026-01 发布的多语言、可控、流式 TTS 模型家族。基于 Qwen3 LLM 系列。

双轨架构Tokenizer 方案核心创新性能
model speech-model open-source codec streaming architecture

Qwen3.5

2026-05-12

Qwen3.5 是阿里巴巴通义千问团队发布的下一代基础模型系列。首个开源权重模型为 Qwen3.5-397B-A17B。

概述架构创新推理优化支持MTP (Multi-Token Prediction)
model architecture open-source inference

Qwen3.5-Omni

2026-04-26

Qwen Team 发布的最新 Omni 模型(arxiv 2604.15804,2026-04-17/21),Qwen-Omni 家族的重大演进。相比 qwen3-omni,在规模、对齐、实时交互上有全面升级。

关键参数架构:Thinker-Talker核心改进延迟指标
model architecture multimodal speech-model training on-policy

SALMONN-omni

2026-04-15

Tsinghua 提出的 Codec-free 全双工语音理解与生成模型。

核心创新能力与其他模型对比关联
model speech-model architecture full-duplex streaming open-source

Seed Realtime Voice

2026-04-09

Seeduplex 的前代模型,即豆包 App 此前使用的端到端语音模型。采用半双工(half-duplex)范式。

关键特征被替代相关页面
model speech-model end-to-end company

Seed-TTS

2026-04-09

字节跳动 Seed 团队于 2024 年 6 月发布的大规模自回归 TTS 模型家族。

基本信息贡献与 Seeduplex 的关系相关页面
model speech-model company

Seeduplex

2026-04-09

字节跳动 Seed 团队于 2026 年 4 月 9 日发布的原生全双工端到端语音大模型,已在中国大陆豆包 App 全量上线,服务上亿用户。

基本信息核心技术突破精准抗干扰(Precision Interference Suppression)动态判停(Dynamic Endpoint Detection)
model speech-model full-duplex end-to-end company

Speech-Omni-Lite

2026-04-15

Speech-Omni-Lite is a portable speech interface designed for vision-language models (VLMs), enabling lightweight speech interaction capabilities. It is referenced in the context of omni-modal-llm as part of the broader ecosystem of multimodal speech-language models. The approach focuses on efficiency and portability, m...

Key FactsRelatedReferences
speech-model multimodal model

Stream-Omni

2026-04-15

ICT/CAS 发布的 text-centric 多模态对齐框架,支持同时视觉理解和语音交互。

核心创新: 差异化模态对齐优势与 Qwen3-Omni 对比关联
model multimodal speech-model architecture streaming open-source

字节跳动 Seed

2026-04-09

字节跳动旗下 AI 研究团队,聚焦大模型基础研究与应用落地。

主要模型成果研究方向相关语音工作相关页面
company lab