The real shift is that embedding models small enough for phones now handle audio and video too, making privacy-first on-device search and RAG practical without cloud calls. Developers building local retrieval or mobile apps can trade a 70–1120 soft-token vision budget against memory. Gains concentrate in code retrieval rather than multilingual text, which only matches the previous version.
Gary Marcus 拆解了 OpenAI 数学成绩背后的门道，说其实靠 Lean 符号验证撑着，顺带回应了关于他的各种争议说法。
Google 开源了个 7.4 亿参数的嵌入模型，一句话就能在手机上搜照片视频和录音，还不用联网，免费商用，做本地搜索的可以去试试。
Google 把 270M 参数的多模态 embedding 模型开源了，手机离线就能搜相册、搜语音，不用传云端，隐私党可以试试。
EmbeddingGemma 2 enables local deployment for high-compliance sectors like healthcare, legal, finance, and government, addressing privacy concerns. For developers in Southeast Asia, this provides a cost-effective alternative to cloud-based multimodal RAG systems when paired with Gemma 4, reducing latency and data transfer costs for applications like local video search in regional languages.
日经统计中国厂商一个月发了 16 个模型，DeepSeek 领跑，还是顶着 Anthropic 警告发的，节奏对比很直观。
OpenAI 放出了 722 篇数学论文，一个内部模型啃了 4000 道题还解决几百个开放问题，数学方向的朋友可以翻翻看
OpenAI 把未发布模型解的约 4000 道数学题解法开了仓库，每题平均烧三小时 Pro 级思考算力，可以去看它的解题过程。
A single averaged score for agent settings can shift how people compare models, since raw benchmark wins often fail to translate into real agent behaviour. Builders selecting a backend for the Hermes Agent get one number to sort by instead of reconciling conflicting leaderboards. Labs gain a new public signal to optimise for, which may influence future model tuning priorities.
GoogleDeepMind的EmbeddingGemma 2来了，首个原生多模态嵌入模型，支持代码图像音频视频统一处理。
The Apache 2.0 license plus modular encoders mean developers can ship private on-device semantic search without per-query API costs, loading only the components an app needs. It matters for anyone building mobile RAG or personal-data search: the 8K context fits roughly 29 images or 5.5 minutes of audio per query, and code-search gains outpace specialist models twice its size.
The key idea is that observability and training share one dataset: the traces showing where an agent fails become the fine-tuning data and the eval set. Teams running agents on specialised documents can replace expensive frontier APIs with a small model tuned to their own workload. The comparison against GPT5.6 Luna on clause-citation accuracy gives a concrete baseline for anyone weighing cost against reliability in production.
OpenAI 要在 GitHub 一次性放出几百道解开的数学难题，WIRED 报道的，前一天晚上论文抢发成潮，挺有意思的动向。
CodeRabbit 拿真实 CVE 隐藏答案考了四个安全智能体，Devin 和 Claude 的分数差距挺意外，写测试方法那部分很实在。
Mistral Large 4 在网络安全基准上击败了 Opus 5.5 和 GPT-6 Astra，安全过滤策略更宽松。
Tavus 新模型 Griffin 实测火了，48% 的人以为在跟真人视频，全双工无剪辑，小动作真实到吓人。
Google 的 Nano Banana 2.1 上 OpenRouter 了，出图速度是 Flash 级，蒙版编辑和主体一致性比上一代强不少，做图的可以试试。
The audit shows that verifier quality can swing agent benchmark pass rates by tens of percentage points, so scores from AutomationBench-era leaderboards may mislead model comparisons. Teams doing independent evaluations should replay runs on the fixed verifiers or borrow the agent-fooling methodology to stress-test their own evals. It is a practical reference point for anyone benchmarking Kimi K3 or similar agents against Zapier-style automation tasks.
欧洲模型在 CyberGym 网络安全基准拿了 82%，压过 MiMo-V2.6-Pro 和 GPT-6 Luna，看来欧洲还没掉队
The speedrun format makes the results verifiable: each solve time and tool call comes from a real run, which is rarer than aggregated benchmark scores. Security researchers and CTF players can compare its tool-calling behaviour against their own agent setups. Compared with benchmark-only announcements from US and Chinese labs, this offers a concrete view of where Mistral's flagship stands.
Mistral 放话 Mistral Large 4 网络安全能力能进全球前列，做安全方向的可以关注下实测
Native multimodal embeddings in one vector remove the need to run separate text and image encoders in retrieval pipelines, and the Apache 2.0 license allows commercial use without restrictions. Developers building RAG, search or on-device applications can pick the 270M text-only variant for low latency or the 740M omni model for mixed-media retrieval.
Co-Director treats long-form video generation as a multi-agent coordination problem rather than a single-model prompt, which is a different architecture from end-to-end video generators like Sora or Veo. Researchers and builders working on agents or video pipelines can read the accompanying blog to study how hierarchy improves narrative consistency. The demo is happening live at COLM 2026, so attendees can see it before the write-ups land.
Mistral 放出 1T 参数的 Mistral Large 4，激活只有 49B，还能超闭源模型做视觉 grounding，权重 10 月底开源，可以关注。
Google 的 Nano Banana 2.1 来了，最多能锁住 4 个角色和 10 个物体不走样，画连环画、做多角度商品图的人可以试试
Google 刚发布的 EmbeddingGemma 2 能把文字、图片、音频、视频都转成向量，vLLM 当天就能跑，做 RAG 检索的可以试试 nightly 版。
小团队 Dots.3-preview 居然在性能和成本上都压过 DeepSeek 新模型，作者重点夸了他们的 RL 训练方案，想追国产模型进展的可以看看。
有人用 LilyPond 让模型写巴赫众赞歌，GPT-6 Astra 第一次写出了经过音，对乐理感兴趣的可以去看看它写的谱子。
Perplexity 放出了开源的决策模型 v1.1-27b，价格砍半到 $0.02/百万 token，还拿了 HF Decision Index 0.3 第一，可以拿来试试。
谷歌新出的 Nano Banana 2.1 能生成干净的中文文字图片，还支持 4K，在 AI Studio 里就能试
The claim reframes the open-source race as an asymmetry in distillation access rather than raw research capability, which changes how model-parity announcements should be read. Builders choosing between GLM-5.2, Beam and Mistral Large 4 get three near-equal open-source options from three different jurisdictions with different licensing and deployment constraints. ASEAN teams weighing Chinese-hosted APIs against Western alternatives can use this tier convergence to compare on price and data residency instead of capability alone.
把一段猫咪发布片丢给 Mistral Large 4、MiMo V2.6 Pro、Opus 5.5 和 GPT-6.1 Sol，看谁重制的 voxel 和 ASCII 版更好玩，结果全贴出来了。
Google 出了 740M 的多模态嵌入模型，文本图片视频都能嵌入，还能配 Gemma 4 离线跑 RAG，权重直接可下。
Google 的 Nano Banana 2.1 来了，圈选局部修图不碰别处，批量出图人物还不变形，做电商和广告素材的可以试试。
ARC Prize 晒出 Grok 4.7 的实测成绩，有意思的是它一代分数涨了，二三代替反而降了，可以看看这套数据对比
Only the latest 2000 items are listed; for older ones see the daily archive
