LLM Models

Date	Model	Size	Context Window	Creator	Ability
2025-05-28	DeepSeek R1 0528
2025-05-20	Gemma3n	8b-e2b, 8b-e4b		Google	Edge, PLE
2025-04-29	Qwen3	0.6b, 1.7b, 4b, 8b, 14b, 30b, 32b, 235b, 30b-a3b, 235b-a22b	40K	Alibaba	MoE, Thinking
2025-04-05	Llama 4	scout 109b-a17b ,marverik 400b-a17b, 2T	1M, 10M	Meta	MoE, Vision
2025-03-26	Qwen2.5-Omni	3B, 7B		Alibaba	text, audio, image, video, speech
2025-03-12	Gemma3	1b, 4b, 12b, 27b	128K	Google DeepMind	Vision
2025-02-26	Wan 2.1			Alibaba	Video
2025-02-24	smollm2	135m, 360m, 1.7b	8K	HuggingFaceTB
2025-01-28	Qwen2.5-VL	3b, 7b, 32b, 72b	125K	Alibaba	Vision
2025-01-28	Qwen2.5	0.5b, 1.5b, 3b, 7b, 14b, 32b, 72b	32K,1M	Alibaba
2025-01-20	DeepSeek R1	1.5b, 7b, 8b, 14b, 32b, 70b, 671b	128K	DeepSeek AI
2024-12-07	Llama 3.3	70B	128K	Meta
2024-10-05	LLaVA	7b, 13b, 34b	4K, 32K
2024-09-25	Llama 3.2	1B, 3B, 11B, 90B	128K	Meta
2024-07-23	Llama 3.1	8B, 70.6B, 405B	128K	Meta
2024-06-27	Gemma 2	9b, 27.2b	8K	Google DeepMind
2024-06-07	Qwen2	0.5b, 1.5b, 7b, 57b (A14b), 72b	32K, 64K, 128K	Alibaba
2024-04-23	Phi-3	3.8b , 7b , 14b	4K, 128K	Microsoft
2024-04-18	Llama 3	8b, 70.6b	8K, 128K	Meta
2024-02-21	Gemma	2b, 7b	8K	Google DeepMind
2023-12-11	Mistral	7b, 46.7b (8x7B MoE)	33K	Mistral AI
2023-07-18	Llama 2	6.7b, 13b, 69b	4K	Meta
2023-02-24	LLaMA	6.7B, 13B, 32.5B, 65.2B	2K	Meta
2020-06-11	GPT-3	175b	2K	OpenAI
2019-02-14	GPT-2	1.5b	1K	OpenAI
2018-06-11	GPT-1	117m	512	OpenAI

Proprietary Models

model	date	notes
GPT-3.5-turbo	2022	4K
GPT-3.5-16k	2022	16K
GPT-3.5	2022	ChatGPT,570GB Text
GPT-4	2023
GPT-4-32k	2023
GPT-4V	2023
GPT-4o	2023
Gemini
Gemini 2.0
Gemini 2.5

*-pt - Pre-Training - 预训练模型
- 在大规模数据集上进行初始训练，学习语言模式和结构。
- 该模型适合作为基础模型，供开发者在特定任务上进行进一步的微调。
*-ft
- Fine-tuned
*-it - Instruction Tuning - 经过指令微调的模型
- 在预训练模型的基础上，进一步针对特定任务或指令进行了微调。
- 此版本更适合直接应用于实际任务，因为它已经针对特定用途进行了优化。
https://ollama.com/library
内存占用计算方式
- 参数x精度
- 目前理想精度是 float16, bfloat16 - 1 个参数占用 16bit
  - 1B -> 2GB
- 量化参数 - 常见量化 int4
  - 1B -> 0.5GB
  - https://huggingface.co/datasets/christopherthompson81/quant_exploration
  - Q4_0 - worse accuracy but higher speed
  - Q4_1 - more accurate but slower
  - q4_2, q4_3 - new generations of q4_0 and q4_1, more accurate
  - https://github.com/ggerganov/llama.cpp/discussions/406
7B - 8GB 内存
13B - 16GB 内存
70B - 32GB/48G 内存
小 context window 适用于 RAG
Context Window
- LLama-3 8B 8K-1M https://ollama.com/library/llama3-gradient
  - 256k context window requires at least 64GB of memory
  - 1M+ context window requires significantly more (100GB+)

按照商业公司分类模型之间关联性高，模型有连续性。虽然会扩展调整各种能力，但是 Base 模型的发展和用到的技术会相对连续。

Leader board
- https://huggingface.co/open-llm-leaderboard
- https://lmarena.ai/
- https://www.vellum.ai/llm-leaderboard
- https://openrouter.ai/rankings
- https://aider.chat/docs/leaderboards/
- https://huggingface.co/models
价格/Pricing
- https://openrouter.ai/models
- https://ai.google.dev/gemini-api/docs/pricing
Visual
- microsoft/Florence-2-large
  - MIT
  - base 0.23B, large 0.77B
  - Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks
阿里云/Alibaba
- QwenLM
  - Qwen
  - QwQ 32B
  - Qwen2 VL
    - Qwen/QVQ-72B-Preview
  - QwenLM/Qwen2.5
    - Qwen2.5 Coder
    - Qwen2.5 VL
    - Qwen2.5 Math
    - Collection Qwen/Qwen2.5-VL
  - QwenLM/Qwen
  - Qwen3
    - 推荐参数
      - Thinking - temperature=0.6, top_p=0.95, top_k=20
      - temperature=0.7, top_p=0.8, top_k=20
      - min_p=0.0
- Wan-Video
  - Wan-Video/Wan2.1
- HumanMLLM/R1-Omni
  - 阿里通义实验室
  - Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning
deepseek
- deepseek-ai/Janus
  - Janus-Series: Unified Multimodal Understanding and Generation Models
- deepseek-ai/DeepSeek-R1
  - MoE, GRPO, MLA, RL, MTP, FP8
- deepseek-ai/DeepSeek-V3
- deepseek-ai/DeepSeek-VL2
  - DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding
- DeepSeek-V2
  - MLA
Google
- SigLIP2
- SigLIP1
- google-deepmind/gemma
  - Apache-2.0, Flax, JAX
  - by Google DeepMind
  - Ultra, Pro, Flash, Nano
  - https://ai.google.dev/gemma/docs/core
  - gemma 3
    - 1B text only, 4, 12, 27B Vision + text. 14T tokens
    - 128K context length further trained from 32K. 1B is 32K.
    - Removed attn softcapping. Replaced with QK norm
    - 5 sliding + 1 global attn
    - 1024 sliding window attention
    - RL - BOND, WARM, WARP
    - 推荐参数: temperature=1.0, top_k=64, top_p=0.95, min_p=0.0
    - ⚠️注意不支持获取对象检测的坐标
    - https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d
    - https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf
    - https://blog.google/technology/developers/gemma-3/
    - https://huggingface.co/blog/gemma3
    - https://blog.roboflow.com/gemma-3/
    - https://docs.unsloth.ai/basics/gemma-3-how-to-run-and-fine-tune
    - Ollama 3 Tools support https://github.com/ollama/ollama/issues/9680
bytedance/字节跳动
- ~~ByteDance-Seed/Seed1.5-VL~~
  - https://huggingface.co/spaces/ByteDance-Seed/Seed1.5-VL
- bytedance-seed/BAGEL
  - ByteDance-Seed/BAGEL-7B-MoT
    - 图片理解、生成、编辑
    - https://huggingface.co/spaces/ByteDance-Seed/BAGEL
    - 要求
      - 1024 × 1024 Image Gen 80GB vRAM
      - 4×16G GPU 能运行
      - e.g. 1 minute on 3xRTX3090, 8 minutes on A100
      - https://github.com/ByteDance-Seed/Bagel/issues/4
    - https://github.com/neverbiasu/ComfyUI-BAGEL
Tencent/腾讯
- https://huggingface.co/tencent
- 混元
- Tencent-Hunyuan/HunyuanVideo-Avatar
  - Image-to-Video
  - https://arxiv.org/html/2505.20156v1
  - https://aivideo.hunyuan.tencent.com/
  - hf tencent/HunyuanVideo-Avatar
- hf tencent/Hunyuan3D-2
  - 2025-01-21
- hf tencent/HunyuanVideo
  - Text-to-Video
  - 2024-10-03
- Tencent/HunyuanVideo-I2V
  - Image-to-Video
  - 720P, vRAM 60GB - 推荐 vRAM 80GB
  - 2025-03-06
  - hf tencent/HunyuanVideo-I2V
    - for diffusers hunyuanvideo-community/HunyuanVideo-I2V
LLaMA based
- Vicuna
haotian-liu/LLaVA
- LLaVA (Large Language and Vision Assistant)
- Vicuna + CLIP
OpenGVLab
- OpenGVLab/InternVL3-8B
command-a
- 主要用于 Agent, 工具调用
- https://cohere.com/blog/command-a
- https://huggingface.co/CohereForAI/c4ai-command-a-03-2025
llama2
- 7B, 13B, 70B
uncensored/abliterated/CensorTune
- Sumandora/remove-refusals-with-transformers
- https://huggingface.co/huihui-ai
- https://huggingface.co/datasets/Guilherme34/uncensor
- https://huggingface.co/models?search=uncensored
- https://erichartford.com/uncensored-models
- https://www.pixiv.net/novel/show.php?id=21039830
  - https://huggingface.co/a686d380/rwkv-5-h-world
microsoft/BitNet
- MIT, C++, Python
- by Microsoft
- HN
vicuna
mistral
mixtral
Flan
Alpaca
GPT4All
Chinese LLaMA
Vigogne (French)
LLaMA
Databricks Dolly 2.0
- https://huggingface.co/databricks/dolly-v2-12b
- https://github.com/databrickslabs/dolly/tree/master/data
https://huggingface.co/stabilityai/stable-diffusion-2
togethercomputer/OpenChatKit
Alpaca
- 基于 LLaMA + 指令训练
FlagAI-Open/FlagAI
hpcaitech/ColossalAI
BlinkDL/ChatRWKV
- ChatGPT like
- RWKV (100% RNN)
nebuly-ai/nebullvm
FMInference/FlexGen
EssayKillerBrain/WriteGPT
- GPT-2
ymcui/Chinese-LLaMA-Alpaca
https://www.promptingguide.ai/zh/models/collection
Releasing 3B and 7B RedPajama-INCITE family of models including base, instruction-tuned & chat models
RedPajama-Data-v2
- https://together.ai/blog/redpajama-data-v2
- https://github.com/togethercomputer/RedPajama-Data
- https://huggingface.co/datasets/togethercomputer/RedPajama-Data-V2
- en, de, fr, es, it
hysts/ControlNet-v1-1
ggml
- ggerganov/ggml
  - MIT, C
.pth - PyTorch
- checklist.chk - MD5
- params.json - {"dim": 4096, "multiple_of": 256, "n_heads": 32, "n_layers": 32, "norm_eps": 1e-06, "vocab_size": -1}
- Saving & Loading Models
https://medium.com/geekculture/list-of-open-sourced-fine-tuned-large-language-models-llm-8d95a2e0dc76
https://erichartford.com/uncensored-models
https://huggingface.co/spaces/facebook/seamless_m4t
https://github.com/LinkSoul-AI/Chinese-Llama-2-7b
Jina AI 8k text embedding
- https://news.ycombinator.com/item?id=38020109
- https://huggingface.co/jinaai/jina-embeddings-v2-base-en
- https://huggingface.co/jinaai/jina-embeddings-v2-small-en
Models
- https://civitai.com/
- https://huggingface.co/models
- https://www.modelscope.cn/models

# AVX = 1 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
grep avx /proc/cpuinfo --color # x86_64

中文

Qwen2
LlamaFamily/Llama-Chinese
UnicomAI/Unichat-llama3-Chinese
- 联通 llama3 微调
https://github.com/datawhalechina/self-llm

Fine-tuning

https://huggingface.co/ValueFX9507/Tifa-Deepsex-14b-CoT-GGUF-Q4

Voice

TTS, STT, ASR, Dialogue, Audio
VITA-MLLM/VITA-Audio
- ASR, TTS, SpokenQA
- https://huggingface.co/spaces/shenyunhang/VITA-Audio
resemble-ai/chatterbox
- MIT, Python
- 开源版本只支持 en
- hf ResembleAI/chatterbox
FunAudioLLM/CosyVoice
- 中文、英文、日文、韩文、中文方言（粤语、四川话、上海话、天津话、武汉话等）
- hf FunAudioLLM/CosyVoice2-0.5B
yl4579/HiFTNet
THUDM/GLM-4-Voice
- 中英语音对话模型
- https://huggingface.co/THUDM/glm-4-voice-tokenizer
- https://huggingface.co/THUDM/glm-4-voice-decoder
  - 基于 CosyVoice 重新训练的支持流式推理的语音解码器，将离散化的语音 Token 转化为连续的语音输出。
https://huggingface.co/datasets/gpt-omni/VoiceAssistant-400K
shivammehta25/Matcha-TTS
modelscope/FunASR
- MIT, Python
- ASR, VAD
nari-labs/dia
- text to dialogue
- 只支持 en
- https://huggingface.co/spaces/nari-labs/Dia-1.6B

MLLM

Multimodal Large Language Model - 多模态大语言模型
结构: 视觉编码器 + 投影器 + 语言模型
Vision Model
- ViT
Language Model
Projector / Vision-Language Adapter
- 将视觉模型提取出的图像特征与语言模型的表示空间对齐
- Cross-Attention Module - 交叉注意力模块

Vision

Document OCR - 文档 OCR
Handwriting OCR - 手写 OCR
Visual QA / Image QA - 图片 QA
Visual Reasoning - 图像推理
Image Classification - 图片分类
Document Understanding - 文档理解
Video Understanding - 视频理解
Object Detection - 对象识别
Object Counting - 对象计数
Agent - 屏幕理解操作
Object Grounding - 物体定位
- 返回 Bounding Box 坐标
- visual grounding poor performance after fine-tuning 2U1/Qwen2-VL-Finetune#77

Qwen2 VL
- factor=28
SmolVLM 256M
- 64 image tokens per 512px image
- https://huggingface.co/spaces/webml-community/smolvlm-realtime-webgpu
  - 500M, SmolVLM
  - https://www.reddit.com/r/LocalLLaMA/comments/1kmi6vl
参考
- https://blog.roboflow.com/multimodal-vision-models/

Coding

All-Hands-AI/OpenHands
block/goose
https://www.swebench.com/
模型
- devstral 24B
  - https://mistral.ai/news/devstral
  - https://ollama.com/library/devstral

Video

整个流程
Flow
harry0703/MoneyPrinterTurbo
- 一键生成高清短视频
- https://huggingface.co/spaces/chaowenguo/avfwae
Lightricks/LTX-Video
- 30 FPS, 1216×704
- text-to-image, image-to-video, keyframe-based animation, video extension, video-to-video
Wan-Video
- https://huggingface.co/spaces/multimodalart/wan2-1-fast

Generation

Text to Image, Video, Audio
Image Inpainting
Image Variation
Text-guided
Upscale, Super Resolution
huggingface/diffusers
- bghira/SimpleTuner
- ostris/ai-toolkit
FLUX.1
- https://playground.bfl.ai/
- collection FLUX.1
- hf black-forest-labs/FLUX.1-dev
  - 12B
- FLUX.1 Kontext
  - distilled variant of Kontext
  - 目前生成的图有点糊
  - https://bfl.ai/models/flux-kontext
  - https://replicate.com/black-forest-labs/flux-kontext-pro
  - https://replicate.com/flux-kontext-apps
  - https://news.ycombinator.com/item?id=44128322
- https://huggingface.co/lodestones/Chroma
https://genai-showdown.specr.net/

问题领域

Prompt adherence（提示词遵循度）
Generation quality（生成质量）
Instructiveness（可指导性）
Consistency of styles, characters, settings, etc.（风格、角色、设置的一致性）
Deliberate and exact intentional posing of characters and set pieces（角色和场景元素的精确姿态和故意摆放）
Compositing different images or layers together（将不同图像或图层组合在一起）
Relighting（重新打光）
Posing built into the model. No ControlNet hacks.（姿态控制内置于模型中，无需ControlNet等“黑科技”）
References built into the model. No IPAdapter, no required character/style LoRAs, etc.（参考功能内置于模型中，无需IPAdapter、角色/风格LoRA等）
Ability to address objects, characters, mannequins, etc. for deletion / insertion.（能够针对物体、角色、人体模型等进行删除/插入操作） Ability to pull sources from across multiple images with or without "innovation" / change to their pixels.（能够从多张图片中提取来源，无论是否对其像素进行“创新”/更改） Fine-tunable (so we can get higher quality and precision)（可微调，以获得更高的质量和精度）

中文​

Fine-tuning​

Voice​

MLLM​

Vision​

Coding​

Video​

Generation​

中文

Fine-tuning

Voice

MLLM

Vision

Coding

Video

Generation