Latest
webAI 开源了 88 亿参数的 TwIL-8B，编码、指令遵循、数学都能打到前沿模型九成左右的成绩，还小到自己的电脑能跑，权重直接放 Hugging Face 上了。
An open-weight 30B MoE with only ~4B active parameters makes speech-to-text with speaker diarisation and translation runnable on modest hardware, undercutting the usual cost of per-language ASR pipelines. Builders working on voice agents or meeting transcription can self-host one checkpoint instead of chaining multiple services. With 40-language coverage, it is a practical candidate for multilingual Southeast Asian speech workloads.
月之暗面把 2.8T 参数的 Kimi K3.1 Pro 权重直接开源了，4-bit 发布省存储，编程基准分数贴近 Claude Opus 5.5，纯知识题却差 GPT-6 一截，许可证对云托管商还有限制。
LMArena 的 Agent Arena 榜单上新了，能看到各智能体模型在不同任务间的取舍对比，选模型前可以先去比一比。
For hour-long agent workloads, the cheapest model is not automatically the right one: Claude Opus 5.5 scores higher but costs twice as much per hour as Gemini 4 Carbon. Teams running long-horizon agents can use these per-hour cost and time figures to match model choice to budget and quality needs. The 55-minute spread among frontier models also suggests task duration, not just accuracy, should enter model selection criteria.
Agent Arena 拆了成本维度看模型：贵任务上 OpenAI 和 Anthropic 不再通吃，Gemini 4 Argon 和小米 MiMo 顶上，选型前值得对号入座。
Cost differences widen precisely on the longest sessions, so the cheapest model for simple tasks is not the cheapest for complex agentic work. Teams running long-horizon agent tasks such as app development can cut bills by re-benchmarking at high task complexity rather than averaging across workloads. The finding also opens the door for Gemini 4 Argon and Xiaomi's MiMo-V2.6 Flash and Pro to challenge OpenAI and Anthropic where budgets are tightest.
Agent Arena 拿 5 家实验室的旗舰模型跑了长任务对比，Claude Opus 5.5 能连续干近 2 小时，Kimi 最快但也最短，差距挺直观。
The result argues that at-scale AI code generation can meet formal verification standards, not just quick prototypes. Mathematics researchers and formal-methods teams gain a machine-checked base they can build on, and it signals AI agents moving from toy proofs to large structured libraries. It also compares favourably with prior partial formalization efforts in scope and coherence.
The zero-data-retention and US-hosting terms matter for teams that couldn't previously route sensitive workloads to a Chinese open-weight model. Developers can now self-host or run K3.1 Pro through Ollama Cloud as an alternative to renting closed APIs from OpenAI or Anthropic. Weight availability on Hugging Face also means inference providers beyond Ollama can deploy it.
Perceptron AI 实测发现标注机器人视频用专用小模型就够，比顶级大模型省 8 倍钱，质量几乎不打折，做机器人的可以看看。
Moonshot AI 出了 Kimi K3.1 Pro，Together AI 上就能直接调用， podcast 里有他们自己人讲细节，想用的话可以先听听。
Moonshot 开放了 Kimi-K3.1-Pro，2.8T 参数只激活 104B，1M 长上下文，BrowseComp 拿了 92.1，权重才 1.56TB，跟 Claude Opus 5.5 掰手腕。
The plug-in design means no retraining of the base model, so teams running open-source models could adopt ZHA by training or downloading a small index module. It matters for anyone serving long-context workloads, where KV-cache reads dominate cost and latency. Compared with MSA, the reported 12.7-point retrieval gain is the concrete benchmark to watch when evaluating sparse attention options.
一个用5300小时视频就练出来的视频识别模型，22个基准平均78.76%，在医疗和遥感任务上还反超了Meta烧几百倍算力的VJEPA，权重已开源。
OpenAI 自己发的 LifeVal 榜单，GPT-6 Astra 比 GPT-5.2 涨了 31 个点，不过任务集没公开，看看就好
腾讯的 Hy4 在 APEX-Agents 上跑赢 DeepSeek V4.1，定价也放出来了，还说要开源权重，开源圈可以盯一下。
字节靠 Seedance 把腾讯、B站变成 API 客户，内容还得在抖音、TikTok 分发，版权又和迪士尼谈妥，商业模式挺值得琢磨。
Artificial Analysis 数了各模型写破折号的频率，GLM-5.3 一条回复平均 7.4 个，Claude Opus 5.5 几乎为零，看你的模型像谁就靠这个。
The Information 消息说 GPT-6.2 这周五就来，还分 Astra、Sol、Cyber 三个方向，可以蹲一下周五的官方消息。
DeepSeek 用华为芯片练 V5，绕开 NVIDIA，2027 年跑全量，这个动向值得盯着看。
中国团队 LiberAI 搞了个巧招：让工人戴 131g 手套干活就能采集机器人训练数据，1 万小时手套数据居然打赢 200 万小时网络视频。
腾讯混元Hy4正式版来了，能看图、操作电脑、跑复杂工作流，还针对Blender和CAD专项训练，API已开放、权重承诺开源。
百度发布的世界模型 Apollo World，能把驾驶场景变成可控仿真来训练 Apollo Go，连传感器数据都实时生成。
Kimi K3.1 Pro 权重开放下载了，2.8 万亿参数，不过想自己跑得准备 8 张 B300，约 60 万美元起步。
LiberAI 用 220 万小时人类视频训练 Liber-0，不用一条机器人演示数据就跑了 60 个任务，做具身智能的可以看看这条路线。
腾讯说自家开源的 Hunyuan Hy4 已经能和最强的开源模型掰手腕了，到底表现如何可以看看这篇报道。
月之暗面把最强的 Kimi K3.1 Pro 放上了 Together AI，100 万上下文还能自己微调，国内模型出海又多一个入口。
Moonshot 新的 Kimi K3.1 Pro 发布当天就能用 vLLM 自己部署了，1M 上下文，Terminal-Bench 4.0 跑 65.7%，自己搭推理服务的话可以试试。
Databricks 第一时间上线了 Moonshot 的 Kimi K3.1 Pro，自家团队已经拿它写代码了，开源模型又多一个能打的选项。
Moonshot 把 Kimi K3.1 Pro 的权重放出来了，想本地跑或者自己微调的可以去 Hugging Face 下载试试。
机器人学操作不用再花钱采机器人演示数据了，直接从人类视频学，60 项任务零样本跑，长时程任务成功率还只有 10.5%，成绩很诚实。
Zero-shot transfer to a physical Franka arm is a harder test than simulation benchmarks, so a ranking here signals which models generalize beyond their training tasks. Robotics teams evaluating foundation models for manipulation can use RoboColosseum's head-to-head setup as a reference when choosing between MolmoAct 2 and GPT-6 Astra.
Offline reference scoring turns DPO from a two-model job into a one-model job, roughly halving memory and removing per-step reference overhead during preference tuning. Teams running preference optimisation on Megatron-scale models can reuse one scoring pass across full hyperparameter sweeps, cutting GPU hours. It matters most for builders fine-tuning large models where a second reference copy was previously the bottleneck.
MIT 的人让 GPT-6 Astra 不训练就改 Qwen 的权重，数学准确率从 7.9% 拉到 48.1%，只动了 0.016% 的参数，有点离谱。
Only the latest 2000 items are listed; for older ones see the daily archive