WEEKLY DIGEST / 2026-W42
本周模型动态
数据采集于 2026-10-18 18:55。新模型、榜单变动、官方发布、论文，分开记录。
首次采集 2026-10-18基线周
新模型
最近 9 天发布或上架
2026-10-17
DeepSeek V4.1 Flash 1017
DeepSeek · 开源 · 输入 $0.2 / 输出 $0.6
2026-10-15
Nvidia Nemotron 3.5 Super VL
NVIDIA · 开源 · 输入 $0.15 / 输出 $0.5 · AA 26.5
2026-10-15
Muse Spark Open
Meta · 开源 · 输入 $1.25 / 输出 $4.25
2026-10-15
FLUX 3 Image Klein
Black Forest Labs
2026-10-13
Gemini 4 Argon
Google · 输入 $2 / 输出 $10 · AA 52.6
2026-10-13
MAI-Image-2.6
Microsoft · 输入 $5 / 输出 $38
2026-10-13
MAI-Image-2.6-Flash
Microsoft · 输入 $1.75 / 输出 $19
2026-10-12
Muse Wonder 1.0 (Meta)
Meta · 输入 $4 / 输出 $20
榜单变动
基线周
本周是基线周。采集器刚建立第一份快照，还没有上一周可以对比。W43 起，这里会显示每个模型的分数涨跌、新入榜和掉榜。
官方动态
厂商 RSS / 博客，最近 8 天
- 10-15Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from MetaHugging Face
- 10-14Developing Gemini 4 Argon responsibly and safelyGoogle DeepMind
- 10-14Gemini 4 Argon for RoboticsGoogle DeepMind
- 10-14Gemini 4 Argon: Built to solve your hardest problemsGoogle DeepMind
- 10-16Welcome Beam: a 501B open model you can own, customize, and scaleHugging Face
- 10-14Run fast, private agents on-device with LFM2.5-24B-A2BHugging Face
本周论文
arXiv cs.CL，评测相关
- 10-15FDE-Gym: The Last Mile of AI Delivery
- 10-15Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?
- 10-15Twice Upon a Time: Complementary Representations for Time-Series Reasoning
- 10-15Training Reasoning Models to Generate Diverse Scientific Ideas
- 10-15Learning Self-Orchestration in Decentralized Agent Swarms
- 10-15MindEval2: Clinically Grounded Patients and Rubrics for Evaluating Language Models in Mental Health Support
- 10-15Grid-Bench: Can LLMs Predict How Discrete Worlds Change?
- 10-15AI Consilium: a Benchmark of LLM Clinical Multidisciplinary Second Opinions
SOURCE HEALTH
本周采集状态
哪个源挂了就写挂了。不用上周数据顶替，也不隐藏失败。
| 采集项 | 状态 | 记录数 | 错误 |
|---|---|---|---|
| models.dev | 正常 | 9048 | |
| openrouter | 正常 | 471 | |
| lmarena:text | 正常 | 416 | |
| lmarena:webdev | 正常 | 144 | |
| lmarena:vision | 正常 | 165 | |
| lmarena:search | 正常 | 42 | |
| lmarena:agent | 正常 | 52 | |
| lmarena:document | 正常 | 44 | |
| lmarena:text_to_image | 正常 | 82 | |
| lmarena:image_edit | 正常 | 59 | |
| lmarena:text_to_video | 正常 | 48 | |
| lmarena:image_to_video | 正常 | 48 | |
| lmarena:video_edit | 正常 | 10 | |
| rss:OpenAI | 正常 | 12 | |
| rss:Google DeepMind | 正常 | 5 | |
| rss:Hugging Face | 正常 | 7 | |
| rss:Google AI | 正常 | 1 | |
| arxiv | 正常 | 8 |
更新节奏：每周一上午自动采集并重建页面。快照存在服务端与本地
data/snapshots/，
每周一份，用于计算涨跌。历史周报从第三周开始归档。