OPEN ARCHITECTURE MAP

MODEL
ATLAS

2021—2022—2023—2024—2025—2026

56MODEL
PROFILES

5MODALITY
TRACKS

INCLUSION ≠ ENDORSEMENT · NO SHARED BACKBONE

CATALOG COVERAGE

20212 PROFILES20222 PROFILES20234 PROFILES20249 PROFILES202511 PROFILES202628 PROFILESOPENEXTENSIBLE

DS V4-ProHCA / CSA DS V3.2DSA Top-2048 GLM 5.2IndexShare ×4 Kimi LinearKDA ×20 GLM-4.7 Flash30B-A3B MiniMax M1Lightning ×70 MiniMax M2.7229B-A10B MiMo V25:1 SWA / Full Nemotron SuperMamba ⇄ Attention TeleChat364L Dense Trunk LongCat Lite14-layer Wide Trunk LongCat 2.0LongCat Indexer Qwen3Thinking Switch Qwen3-NextGated DeltaNet ×36 DeepSeek V2MLA Llama 3.1Dense Transformer Mixtral8 Experts Top-2 Mamba 2SSD

Kimi K2.7MoonViT 27L Kimi K2.5MoonViT 400M GLM OCRCogViT 24L MiniMax M3MSA Step 3.73:1 SWA / Full MiniCPM-o 4.5Full Duplex Nemotron OmniMamba ×23 Qwen 3.6Gated DeltaNet ×30 LongCat NextAny-to-Any Tokens Qwen2.5-VLDynamic Resolution LLaVA 1.5CLIP ViT-L/14 Janus-ProDecoupled Visual Encoding CLIPDual Encoder SAM 2Hiera Encoder

Qwen3 ASR24L Audio Encoder Qwen3 TTS12.5Hz Codec VibeVoice ASR60-min Single Pass Voxtral TTS3L Acoustic Flow KittenTTS15M Parameters MOSS TTS NanoGlobal-Local Transformer PersonaPlexFull Duplex Diffusion ASRMasked Diffusion WhisperEncoder–Decoder SenseVoiceSAN-M Encoder VITSConditional VAE F5-TTSConditional Flow Matching CosyVoice 3Speech Token LLM

FLUX.2 devDouble ×8 GLM ImageAR → DiT SD 1.5Latent Diffusion SDXLBase + Refiner DiT-XL/228L Transformer SD 3.5MMDiT FLUX.119 Double + 38 Single SANA 1.532× DC-AE

LTX-214B Video + 5B Audio Wan2.140L Video DiT HunyuanVideo20 Double + 40 Single

SWIPE TO EXPLORE THE MODEL FIELD

56 RESULTS

FILTERORGANIZATION

ARCHITECTURE TIMELINE

沿着模型演进，定位每一次结构选择

年份是主索引；同一时期的模型按研究顺序聚合。卡片只保留入口决策所需信息，完整 Graph、FLOPs、Tensor 与点评进入模型 MRI 后查看。

LONG CONTEXT · AGENTIC · MULTIMODAL

18QwenMULTIMODAL

2026.04 · HYBRID VLM

Qwen3.6-35B-A3B

30 层 Gated DeltaNet 与 10 层全局 Attention 交替，35.95B 容量只激活约 3B。

MODEL SIZE35.95Btotal params

ACTIVE3Bper token

CONTEXT262Ktokens

Gated DeltaNet ×30Full Attention ×10256-way MoENative Vision

71.92GB · BF16 · Apache-2.0OPEN MODEL MRI

2026.01 · STREAMING ASR

Qwen3-ASR-1.7B

24 层声学编码器接 28 层 Qwen3 解码器，统一支持流式与离线的 52 种语言/方言。

MODEL SIZE2.35Btotal params

ACTIVE2.35Bruntime path

DEPTH24 + 28blocks

CONTEXT65Kstream / unit

24L Audio Encoder28L Qwen3 DecoderStreaming / Offline52 Languages

4.70GB · BF16 · Apache-2.0OPEN MODEL MRI

2026.01 · 12HZ TTS

Qwen3-TTS-12Hz-1.7B

12.5Hz 多码本 audio tokenizer 配合 28 层 Talker 与 5 层 Code Predictor，实现约 97ms 首包。

MODEL SIZE1.93Btotal params

ACTIVE1.93Bruntime path

DEPTH28 + 5blocks

CONTEXT32Kstream / unit

12.5Hz CodecDual-Track Streaming28L Talker5L Code Predictor

4.54GB · BF16 · Apache-2.0OPEN MODEL MRI

21MicrosoftAUDIO

2026.01 · 60-MIN ASR

VibeVoice-ASR

ConvNeXt 式双声学 tokenizer 与 28 层文本解码器单次处理最长 60 分钟，并输出说话人和时间。

MODEL SIZE8.33Btotal params

ACTIVE8.33Bruntime path

DEPTH2×7 + 28blocks

CONTEXT131Kstream / unit

60-min Single PassDual Audio TokenizerSpeaker DiarizationHotwords

16.67GB · BF16 · MITOPEN MODEL MRI

22Mistral AIAUDIO

2026.03 · FLOW TTS

Voxtral-4B-TTS-2603

26 层 3.4B 文本主干驱动 3 层声学 Flow Transformer，以 8 个 Euler 步生成 24kHz 语音。

MODEL SIZE4.0Btotal params

ACTIVE4.0Bruntime path

DEPTH26 + 3blocks

CONTEXT128Kstream / unit

3L Acoustic Flow8 Euler Steps20 Voices9 Languages

8.04GB · BF16 · CC-BY-NC-4.0OPEN MODEL MRI

23KittenMLAUDIO

2026.02 · EDGE TTS

KittenTTS Nano 0.8

约 15M 参数、INT8 ONNX 仅 27.7MB，面向无 GPU CPU 语音合成；官方未声明逐层拓扑。

MODEL SIZE15Mtotal params

ACTIVE15Mruntime path

DEPTHONNX DAGblocks

CONTEXTsentencestream / unit

15M ParametersINT8 ONNXCPU Inference8 Voices

27.7MB · INT8 · Apache-2.0OPEN MODEL MRI

24OpenMOSSAUDIO

2026.04 · 100M TTS

MOSS-TTS-Nano-100M

12 层 768 维 global-local Transformer 自回归生成 16 组音频码本，在四核 CPU 上流式合成 48kHz 立体声。

MODEL SIZE100Mtotal params

ACTIVE100Mruntime path

DEPTH12 + localblocks

CONTEXT32Kstream / unit

Global-Local Transformer16 Codebooks48kHz Stereo20 Languages

706.97MB repo · FP32 · Apache-2.0OPEN MODEL MRI

2026.03 · DUPLEX SPEECH

PersonaPlex-7B-v1

Moshi/Helium 式 7B 主干同时读取和生成 8 路 Mimi 音频码本，以文本角色提示和声音提示控制人格。

MODEL SIZE8.37Btotal params

ACTIVE8.37Bruntime path

CONTEXTstreamstream / unit

Full DuplexMimi 8 CodebooksText Role PromptVoice Conditioning

17.14GB · BF16 · NVIDIA Open ModelOPEN MODEL MRI

26LightricksVIDEO GEN

2026.03 · A/V GENERATION

LTX-2 19B

48 个非对称双流 DiT block 由 14B 视频分支与 5B 音频分支组成，在同一步中联合生成画面和声音。

MODEL SIZE18.88Btotal params

ACTIVE18.88Bper denoise step

CONTEXTvideo latentlatent grid

14B Video + 5B Audio48 Dual-Stream BlocksBidirectional A/VDistilled 8 Steps

43.3GB BF16 · BF16 / FP8 · LTX CommunityOPEN MODEL MRI

2026.03 · HYBRID MOE

Nemotron 3 Super 120B-A12B

88 层 Mamba-Transformer 混合主干配 512 路 LatentMoE、top-22 与 2 层 MTP，原生上下文达 1M。

MODEL SIZE123.6Btotal params

ACTIVE12.7Bper token

DEPTH88 + 2blocks

CONTEXT1Mtokens

Mamba ⇄ Attention512 Latent ExpertsTop-22MTP ×2

247.25GB · BF16 · NVIDIA Open ModelOPEN MODEL MRI

2026 · DENSE REASONING

TeleChat3-36B-Thinking

64 层、6,144 hidden 的标准 dense GQA 主干，结构不追求新奇，提供一个 36B 深宽平衡的基准。

MODEL SIZE36.24Btotal params

ACTIVE36.24Bper token

CONTEXT32Ktokens

64L Dense TrunkGQA 48/8YaRN ×4Thinking Tokens

72.48GB · BF16 · TeleChat CommunityOPEN MODEL MRI

29InterfazeAUDIO

2026.07 · DIFFUSION ASR

Diffusion-Gemma-ASR Small

冻结 Whisper-small 与 DiffusionGemma-26B-A4B，只训练约 42M adapter，在固定 256-token 画布上并行去噪转写。

MODEL SIZE26B + 42Mtotal params

ACTIVE≈4Bruntime path

DEPTHWhisper + 48blocks

CONTEXT256 canvasstream / unit

Masked Diffusion42M Trainable AdapterWhisper Encoder26B-A4B Backbone

191.08MB adapter · BF16 adapter · Apache-2.0OPEN MODEL MRI