Chinese contentPerplexity 发布 pplx-embed-v2-late 嵌入模型，0.6B 与 9B 两个规格
Perplexity 开源了 0.6B 和 9B 两款多模态嵌入模型，0.6B 就能打大模型，还支持 9B 建库、0.6B 查询的省资源玩法，做检索的可以试试。
Hugging Face
United States·New York
The hub for open models and datasets
What they do
Founded in 2016 and headquartered in New York, Hugging Face runs the largest hosting platform for open models and datasets and maintains open-source libraries such as Transformers.
Latest coverage
Stories in the feed that mention this companyChinese content黄仁勋在 GTC Berlin 2026 收官演示 Hermes Agent：一句话装好并训练机器鸭
老黄在 GTC Berlin 用一条提示词装好 Hermes，再让它训练几千只模拟机器鸭跑障碍赛，开源本地+云端模型混搭玩法很值得看。
AWS shows three ways to train Qwen3 on Trainium with torchtitan
The notable point is that one config setup now covers three different training backends on Trainium, so teams can switch between them without rewriting launch scripts or breaking checkpoint compatibility. Engineers choosing between NVIDIA GPUs and AWS silicon can use this to evaluate Trainium for fine-tuning open-source models like Qwen3. The guide also shows how NKI MoE kernels fit into the same workflow.
Chinese contentRecursive 发布 Flint：本地编码模型，MacBook Pro 上跑出 125+ tokens/sec
Richard Socher 团队出了个本地编码模型 Flint，MacBook 上每秒能跑 125+ tokens，权重开放，不用联网不用按量付费，写代码的可以试试。
Chinese contentTorchNeuron 性能剖析教程：在 AWS Trainium 上做系统级 profiling
在 Trainium 上跑 PyTorch 但不知道慢在哪的朋友，这篇教你怎么用 Neuron Profiler 找瓶颈，很实用。
DeepSpeed adds AutoSP, Muon support and auto parallelism at PyTorchCon
AutoTP and AutoEP applied to Hugging Face models mean teams can distribute training without hand-tuning tensor or expert parallelism, lowering the barrier for mid-size labs. Practitioners running multi-vendor accelerator fleets gain a portable optimization stack outside vendor-locked toolchains. As DeepSpeed is open source, the features are immediately available to ASEAN startups building training infrastructure on limited budgets.
TorchTPU brings native PyTorch support for Google TPUs
PyTorch developers can now target TPU hardware through the same APIs they already use, reducing the migration cost that historically kept most teams on NVIDIA GPUs. Teams running training or inference pipelines with TorchTitan, vLLM, or SGLang can evaluate TPUs without rewriting code. The use of agents for kernel generation and sharding also hints at how optimization work itself may be automated.
Chinese contentNVIDIA 发布 SimReady Central：面向机器人训练的仿真级 3D 资产开源库
NVIDIA 在 Hugging Face 开源了一批做机器人仿真的 3D 资产，物理参数都调好了，训练机器人可以直接拿来用，省去自己建模的时间。
Chinese content蚂蚁百灵开源 Ming-3.0 语音系列五款模型，覆盖 TTS、音频理解与 ASR
蚂蚁一口气开源了五款语音模型，TTS 还带导演模式，能指定哪句重读、情绪怎么变，直接下载或调 API 都行。
IT之家Source
Chinese content研究分析 5398 个仓库：开发者实际只用到一小部分 LLM 推理配置项
三个后端、五千多个仓库被扒了一遍配置，想知道大家平时真的调哪些推理参数、哪些基本没人碰，看这篇就够了。
Chinese contentOrnith-1.5-122B 开源发布，多数评分超过 Qwen3.8-Flash-Next
一个 MoE 开源模型，122B 只激活 10B，跑起来不算重，作者称评分多数超过 Qwen3.8-Flash-Next，可以直接下载试试。
