[GPT-3]:Language Models are Few-Shot Learners:2020.6(⭐⭐⭐⭐)
- 开山鼻祖,直接促成了后续ChatGPT的诞生
- 175B参数,上下文学习
[GPT-4]:GPT-4 Technical Report:2023.3 (⭐⭐⭐⭐⭐)
- 对ChatGPT技术的全面展示;在此之后OpenAI开始转向商业化,很难再有如此的细节披露了
The Claude 3 Model Family: Opus, Sonnet, Haiku:2024.3
Introducing Claude 4:2025.5
Introducing Claude 3.5 Sonnet:2024.6
- 确立了Claude在当时Coding模型领域的地位
Introducing Claude Sonnet 4.5: 2025.9
Introducing Claude Opus 4.5:2025.11
- 以上基本没有讲技术细节,就是把benchmark的结果贴了出来
- 可解释性, 安全是Anthropic的关注重点并且在对外宣传中一直被反复强调
Anthropic还有两个博客,会发布一些重要/不重要的技术细节
https://www.anthropic.com/research
https://www.anthropic.com/engineering
Gemini: A Family of Highly Capable Multimodal Models
- 平平无奇
Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
Gemini 3 链接
- 扭转了Google在大模型领域的颓势,并给OpenAI造成了压力
Gemini 3.1-Pro Model Card: 链接
Gemma: Open Models Based on Gemini Research and Technology
Gemma 2: Improving Open Language Models at a Practical Size
Gemma 3 Technical Report
Gemma4 Report Unofficial Introduction
The Llama 3 Herd of Models: 2024.7 ⭐⭐⭐
- 开源领域重要模型
- 后续的Llama 4 陷入争议,连同arxiv被撤稿
Introducing Muse Spark: Scaling Towards Personal Superintelligence:2026.4
PENDING
Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving: 2025.9
Kimi k1.5: Scaling Reinforcement Learning with LLMs: 2025.1 https://arxiv.org/abs/2501.12599
Kimi K2: Open Agentic Intelligence: 2025.7 https://arxiv.org/abs/2507.20534
- 最早明确使用Muon 优化器的Kimi模型
Kimi K2.5: Visual Agentic Intelligence: 2026.2 https://arxiv.org/pdf/2602.02276 ⭐⭐⭐⭐
- 非常简洁,用的是主流的技术;强调了token-efficiency。但是细节不是很多,几乎是一笔带过。而在Agentic Infra上面似乎也没有讲太多东西。
- Agent Swarms可能是独有或单独提到的技术
Kimi K3: ⭐⭐⭐⭐⭐ , 2026.7
- 架构大杂烩,加入了2025就提出来的KDA,以及全新的AttnRes;此外还有GatedMLA,Stable LatentMoE,MoonViT-V2
- 重要意义在于追上了当时最先进闭源大模型Fable 5的水准,并且在前端代码方面展示了自己独一无二的实力
- 是原生多模态的大模型,所谓原生,是指从训练之初就用到了多个模态的数据在同一个主架构里进行联合训练,但是这仍然会用到一个视觉Encoder。
- Post-Train:Multi-Teacher Distillation;GRM升级为Agentic GRM(这个事实上可能会导致更长的耗时);个人助理数据构造,Gmail, Notion, Slack, and Canvas作为低成本的模拟器
- Infra:Virtual Stages(应该是受到了网络负载均衡的启发),MoE balanced等多个优化
- Infra 推理:略(**吐槽:**即使做了如此多的优化Kimi K3仍然是非常慢的一个模型)
GLM: General Language Model Pretraining with Autoregressive Blank Infilling: 2022.5
Chatglm-rlhf: Practices of aligning large language models with human feedback: 2024.4
Chatglm-math: Improving math problem-solving in large language models with a self-critique pipeline: 2024.4
GLM-130B: An open bilingual pre-trained model: 2022.10
ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools: 2024.6
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models: 2025.8
GLM-5: from Vibe Coding to Agentic Engineering: 2026.2
-
重点放在了异步RL;
-
GLM-5 On-Policy Cross-Stage Distillation
中规中矩,没有太多技术创新;
Thinking的三种模式似乎只是格式的变化,而且又增加了认知的成本;
后训练:IcePop 技巧缓解训推不一致性;
DSA=Indexer;
ORM, PRM, GRM的三重奖励;
**Code环境:**Repo-Launch
computationally prohibitive 这个单词太难受了,能不能改成computationally unaffordable。
异步时,生成一条轨迹的过程中模型可能已更新多次,导致无法精确追踪行为策略的概率 。存储多个旧检查点又不现实。优化1:直接复用生成轨迹时实际使用的 log 概率。优化2:超出clip区间的不参与梯度更新;
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model,2024.5
DeepSeek-Prover: Advancing Theorem Proving in LLMs through Large-Scale Synthetic Data,2024.5
DeepSeek-V3 Technical Report,2024.12
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models,2024.2
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,2025.1,⭐⭐⭐⭐⭐
DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition,2025.4
DeepSeek-V3.2, DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models, 2025.12 ⭐⭐⭐
- Unbiased K3 Estimator
Deepseek-V4(Flash/Pro): 2026.4 ⭐⭐⭐⭐
-
继承了Kimi Muon优化器的衣钵;用到了自己创新的mHC;
-
Infra 味非常浓,介绍了EP Scheme如何通过通信和计算的流水线化降低延迟,DeepGEMM的使用;
-
提到了Ascend芯片的适配问题;
-
在算子层面,提到了新锐TileLang,这也是北大系的重要贡献;提到了一些虽然看起来像魔法但其实又在情理之中(不代表容易)的技巧;例如紧凑的代码表示,SMT-Solver-Assisted技巧;
-
批次不变性:Batch Invariance,可能是受到了Thinking Machine Lab的影响,希望去掉训练的随机性;这么做的好处是为了方便debug;
-
细粒度的checkpoint机制的微分(本来要用checkpoint减少显存占用,但是checkpoint的粒度太大,计算量增加,所以要手写减少计算量,但是手写又不能自动微分,开发量太大,所以写了个自动微分的方法);
-
CSA, HCA(跟NSA,DSA的关系);
-
XML天然不需要转义;
-
loss-spike的解决方法:nticipatory Routing(预期路由) 和 SwiGLU Clamping(SwiGLU 钳位)
-
agentic架构的变化;
-
**Code环境:**microVM:Firecracker
Qwen2-VL, Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution, 博客 2024.9
Qwen2.5-VL, Qwen2.5-VL Technical Report,2025.2
Qwen3-VL, Qwen3-VL Technical Report,2025.12
Qwen2-Audio, Qwen2-Audio Technical Report,2024.7
Qwen2.5-Omni, Qwen2.5-Omni Technical Report,2025.3
Qwen3-Omni, Qwen3-Omni Technical Report,2025.9
Qwen2.5-Coder, Qwen2.5-Coder Technical Report,2024.9
Qwen3-Coder-Next, Qwen3-Coder-Next Technical Report,2026.2
Qwen2-Math, Qwen2-Math Technical Report,2024.8
Qwen2.5-Math, Qwen2.5-Math Technical Report,2024.9
Qwen3-Embedding, Qwen3 Embedding Technical Report,2025.6
Qwen3-VL-Embedding, Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking,2026.1
Qwen-Image, Qwen-Image Technical Report,2025.8
Qwen3-ASR, Qwen3-ASR Technical Report,2026.1
Qwen3-TTS, Qwen3-TTS Technical Report,2026.1
Qwen3Guard, Qwen3Guard Technical Report,2025.10
QwenLong, QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning,2025.5
Qwen, Qwen Technical Report,2023.9
Qwen2, Qwen2 Technical Report,2024.7
Qwen2.5, Qwen2.5 Technical Report,2024.12
Qwen3, Qwen3 Technical Report,⭐⭐⭐⭐2025.5
- 国内开源第一梯队;超级全家桶;值得一看是Qwen3-VL, Qwen3 Technical Report
Seed-Prover, Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving,2025.7
Seed-Prover 1.5, Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience,2025.12
Seed diffusion: A large-scale diffusion language model with high-speed inference
Seedance 2.0: Advancing Video Generation for World Complexity
- 视频生成领域的第一(2026年)
Hunyuan-3: Hy3
参考INT2的实现方法:
Hy3 模型:https://huggingface.co/tencent/Hy3
低比特GGUF模型:https://huggingface.co/AngelSlim/Hy3-GGUF
llama.cpp patch : https://huggingface.co/AngelSlim/Hy3-GGUF/tree/main/patches
https://huggingface.co/AngelSlim/Hy3-GGUF/blob/main/patches/01-hyv3-arch.patch
2026年 6 月 16 日,阿里正式推出通义千问首款全系列完整具身智能模型 Qwen-Robot,该系列由三款模型构成:分别是依托超 38100 小时开源操作数据完成训练的 VLA 操作模型 Qwen-RobotManip、面向移动场景的 VLN 导航模型 Qwen-RobotNav,以及以自然语言为动作交互接口、融合二十余种机器人本体开展联合训练的世界模型 Qwen-RobotWorld。三款模型既支持单独部署使用,也能够联动协同运行。(来源:机器人全球资讯)
这三篇论文在具身智能业界算得上一股清流,信息密度很大,内容富有借鉴意义
其实从Qwen3-VL技术报告可以看出通义其实内部对多模态已经有相当好的技术积累。所以单纯从VL出发推进的具身智能技术路线有这样的水平并不意外。
Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System arXiv: 2606.18112,2026.6
Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models arXiv: 2606.17846,2026.6
Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation arXiv: 2606.17030,2026.6
三篇论文的关系:
| 论文 | 解决的问题 | 对应能力 |
|---|---|---|
| Qwen-RobotNav | 机器人如何在环境中移动 | 空间导航 |
| Qwen-RobotManip | 机器人如何操作物体 | 物理交互 |
| Qwen-RobotWorld | 行动后世界会如何变化 | 世界预测 |
模型架构(GDN, DSA, NSA,MoE)
RL 异步框架
大规模预训练
多模态
Agentic技术
数据
