LLM Awesome
Open Weights / Transformer
| Date | Model Series | Size | Context Window | Creator | Notes |
|---|---|---|---|---|---|
| 2026-04-07 | GLM 5.1 | 754B MoE | 200K | Zhipu AI | 大规模稀疏 MoE |
| 2026-04-02 | Gemma 4 | E2B / E4B / 26B / 31B | 128K-256K | 原生多模态、MoE 变体 | |
| 2026-03-12 | MiniMax 2.7 | 230B，10B active | 200K | MiniMax | Sparse MoE |
| 2026-02-16 | Qwen 3.5 | 0.8B-397B，17B active | 256K-1M | Alibaba | 模型家族、MoE |
| 2026-02-11 | GLM 5 | 744B，40B active | 200K | Zhipu AI | Sparse MoE |
| 2025-12-23 | GLM 4.7 | 358B | Zhipu | MoE、文本、代码、多语言 | |
| 2025-12-21 | DeepSeek V3.2 | 685B | DeepSeek | ||
| 2025-12-20 | MiniMax M2.1 | 228B | MiniMaxAI | 文本生成、对话 | |
| 2025-12-16 | MiMo-V2-Flash | 310B | Xiaomi | 文本生成、对话 | |
| 2025-09-29 | DeepSeek-V3.2-Exp | 685B | DeepSeek | 文本生成、对话 | |
| 2025-09-15 | Qwen3-Next | 80B | Alibaba | ||
| 2025-08-20 | Seed-OSS-36B-Instruct | 36B | ByteDance-Seed | 文本生成、对话 | |
| 2025-08-06 | GPT OSS | 20B / 120B | 128K | OpenAI | Reasoning、Tools |
| 2025-07-28 | GLM 4.5 | 355-32A / Air 106-12A | 128K | Zhipu | Reasoning、多语言 |
| 2025-07-23 | Qwen3 2507 | 30B-A3B / 235B-A22B / Coder | 256K / Yarn 1M | Alibaba | Reasoning、Coding |
| 2025-07-11 | Kimi K2 | 1T-A32B | 128K | Moonshot AI | MoE |
| 2025-06-11 | Magistral | Small 24B | 39K | Mistral AI | Reasoning、多语言 |
| 2025-06-07 | Comma v0.1 | 7B | EleutherAI | Full OSS、English | |
| 2025-06-05 | Qwen3-Embedding | 0.6B / 4B / 8B | 32K | Alibaba | Embedding、Reranking、多语言 |
| 2025-04-29 | Qwen3 | 0.6B-235B | 40K | Alibaba | MoE、Reasoning |
| 2025-04-05 | Llama 4 | Scout / Maverick | 1M / 10M | Meta | MoE、Vision |
| 2025-03-26 | Qwen2.5-Omni | 3B / 7B | Alibaba | Text、Audio、Image、Video、Speech | |
| 2025-03-12 | Gemma 3 | 1B / 4B / 12B / 27B | 128K | Google DeepMind | Vision |
| 2025-01-20 | DeepSeek R1 | 1.5B-671B | 128K | DeepSeek AI | Reasoning |
| 2024-12-07 | Llama 3.3 | 70B | 128K | Meta | |
| 2024-12 | Phi-4 | 14B | 128K | Microsoft | Reasoning、多模态 |
| 2024-09-25 | Llama 3.2 | 1B / 3B / 11B / 90B | 128K | Meta | |
| 2024-07-23 | Llama 3.1 | 8B / 70.6B / 405B | 128K | Meta | |
| 2024-06-07 | Qwen2 | 0.5B-72B | 32K-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 | |
| 2023-12-11 | Mistral | 7B / 46.7B | 33K | Mistral AI | Dense / MoE |
| 2023-07-18 | Llama 2 | 7B / 13B / 70B | 4K | 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
| release | model | output | input price | author | notes |
|---|---|---|---|---|---|
| 2026-03-12 | MiniMax M2.7 | $1.20/1M | $0.30/1M | MiniMax | 205K、Sparse MoE API |
| 2026-02-11 | GLM 5 | $2.30/1M | $1.00/1M | Zhipu AI | 200K |
| 2025-06-17 | Gemini 2.5 Pro | $10.00/1M | $1.25/1M | 1M | |
| 2025-06 | Gemini 2.5 Flash | $2.50/1M | $0.30/1M | Audio $1.00/1M | |
| 2025-05-22 | Claude 4 Opus | $15/1M | $3/1M | Anthropic | 200K |
| 2025-05-22 | Claude 4 Sonnet | $75/1M | $15/1M | Anthropic | 200K |
| 2025-04-14 | GPT-4.1 / mini / nano | OpenAI | |||
| 2025-03-25 | Gemini 2.0 Pro | 2M | |||
| 2025-01-10 | o3 / o3-mini | OpenAI | Reasoning | ||
| 2024-05-13 | GPT-4o | OpenAI | Text、Audio、Image | ||
| 2024-03-04 | Claude 3 Haiku / Sonnet / Opus | Anthropic | 200K | ||
| 2024-02-15 | Gemini 1.5 Pro | 1M context | |||
| 2023-11-06 | GPT-4 Turbo / GPT-4V | OpenAI | 128K、Vision | ||
| 2023-03-14 | GPT-4 | OpenAI | 8K / 32K、Image |
Architecture Notes
| date | model | parameters | context | attention | layers | experts | data type |
|---|---|---|---|---|---|---|---|
| 2025-08-05 | gpt-oss-120b | 117B，5.1B active | 128K | GQA + Sparse | 36 | 128 / 4 | MXFP4 / bf16 |
| 2025-08-05 | gpt-oss-20b | 21B，3.6B active | 128K | GQA + Sparse | 24 | 32 / 4 | MXFP4 / bf16 |
| 2025-05 | Qwen3-30B-A3B | 30.5B，3.3B active | 256K | GQA | 48 | 128 / 8 | bf16 |
| 2025-05 | Qwen3-32B | 32.8B | 128K | GQA | 64 | Dense | bf16 |
| 2024-12 | DeepSeek V3 | 671B，37B active | 128K | MLA | 61 | 256 routed + 1 shared | fp8 / bf16 |
| 2019-02-14 | GPT-2 1.5B | 1.542B | 1024 | Multi-Head | 48 | Dense | fp16 |
中文与开源生态
- QwenLM/Qwen
- deepseek-ai/DeepSeek-R1
- MoE、GRPO、MLA、RL、MTP、FP8。
- meta-llama
- haotian-liu/LLaVA
- Vicuna + CLIP 的视觉语言助手。
- ymcui/Chinese-LLaMA-Alpaca
- FlagAI-Open/FlagAI
- BlinkDL/ChatRWKV
- ChatGPT-like，RWKV RNN 架构。
- databricks/dolly-v2-12b
- togethercomputer/OpenChatKit
- Alpaca
- ggml-org/ggml
- C/C++ 推理和张量计算生态。
Abliterated
- UGI - Uncensored General Intelligence。
- Norm-Preserving Biprojected Abliteration。
- Abliteration 介绍
- UGI Leaderboard
- ArliAI
- Sumandora/remove-refusals-with-transformers
推理、量化与内存
- 理想精度：
float16/bfloat16，约 1B 参数占 2GB 权重内存。 - 常见
int4量化约 1B 参数占 0.5GB，实际还要加 KV cache、运行时和临时 buffer。 - 小 context window 适合 RAG；上下文越长，KV cache 和显存需求越高。
Q4_0、Q4_1、Q4_K等量化格式需要结合质量、速度和 runtime 评估。- 经验值：7B 约 8GB、13B 约 16GB、70B 约 32-48GB，实际取决于精度和上下文。
- llama.cpp
- Hugging Face Models
- ModelScope Models
- Ollama Library
Leaderboard 与评估
- Open LLM Leaderboard
- LM Arena
- LiveBench
- OpenRouter Rankings
- Aider Leaderboards
- BFCL Leaderboard
- VLMEvalKit
- SWE-bench