搜索“GitHub”

formant and efficient. The major features for production-grade agentic workloads are already supported. More performance optimizations and advanced features are under way – open a Github issue to request collaboration on yours! Across the two engines, most of the popular open model families are supported, including Qwen, DeepSeek, GLM, Kimi, GPT-OSS, Gemma4 and more. The team is investing in making more models available to run natively on TPUs, including Day 0 support for new open models. We als

匹配字段正文

ek再放大招，全面开源自研推理框架DeepInfer。 一个集群一天服务155万亿Token，DeepSeek V4.1 Flash高吞吐、低延迟的大规模推理，靠的就是它。 一起放出的，还有生产级Token编解码组件deepseek-recipe。 英伟达SM100 GPU、华为昇腾950 NPU，都在支持名单上。 DeepSeek还在知乎发了技术长文，又在GitHub放出一份部署问答，把这套系统从架构设计到部署细节，一五一十讲了出来。 按官方公布的线上统计，平均每个请求输入77K Token，首字平均只要400毫秒就能出来。生成阶段每秒输出230个Token，写代码时能冲到300个。 平均96%的输入，都能直接命中缓存，根本不用重算。 按谷价算，平均每张卡每小时的收入是38美元。 现如今，全世界都可以看到，DeepSeek究竟是如何让大模型高效接活儿的。 DeepSeek研究员Deli Chen更是放了一句狠话。 开源与闭源。光明与黑暗。两种未来、人类两种命运之间的决战。他们想打多久，我们就打多久。直到取得完全胜利。 评论区下，满屏都是开发者的狂欢。 正如网友所言，模型开源已是「基操

匹配字段正文

ts with competing priorities sharing the same infrastructure, and upgrades that do not interrupt operations. What Version 1.0 Adds The project’s v1.0.0 release notes, published on GitHub on October 20, 2026, with signed and verified tags, lead with multi-tenant quality of service: priority- and fairness-aware flow control with controlled load shedding, including a probabilistic admitter, program-aware fairness, and turn-priority scoring, so latency-sensitive requests and bursty agent workloads c

匹配字段正文

25%，从 384 MiB 降至 287 MiB，并为采用更快矩阵乘法和原生 FP8 矩阵乘法的 Metal 4.1 着色器奠定基础。 总结 PyTorch 2.15 将 FP8 带到 Mac，为稠密和 MoE 模型带来更快的 LLM 解码，在 GPU 上提供更多线性代数运算，并将一长串运算从 MPSGraph 迁移到专用 Metal 内核。快来试试候选发布版吧！ 我们的下一步 更多 MPS：2.16 将完全在 Metal 上运行注意力，我们也会继续将运算从 MPSGraph 迁移出去。 torch.compile：我们正在努力消除图中断，让更多模型能够端到端编译。 你的想法：我们还能在 PyTorch 中做些什么？如果有哪里慢、缺失或损坏，无论是在 MPS 还是其他地方，请在 GitHub 上开一个 issue 告诉我们。 非常感谢 Nikita Shulga、Kurt Mohler 和 Aaron Gokaslan 审阅了本项工作的大部分内容，也感谢所有在本版本中为 MPS 做出贡献的人。 博客更多文章 tokenizers v1：编码、解码与扩展，实测 解锁连续批处理中的异步性

匹配字段正文

馈和贡献塑造了我们的多项重要功能。感谢所有贡献者帮助 AnyLanguageModel 对应用开发者更加实用。 下一步 语言模型已被证明是为产品注入智能的强大工具。它们也是开发者可以使用的众多模型类型之一。我们很高兴将同样的方法扩展到更多能力：熟悉的 Swift API，以及可在设备上运行的多种模型选择。即将推出：能够决策的模型，以及能够聆听的模型。 链接 GitHub 上的 AnyLanguageModel GitHub 上的 swift-huggingface GitHub 上的 swift-xet 隆重推出 AnyLanguageModel 隆重推出 swift-huggingface 本文提到的模型 4 LiquidAI/LFM2.5-2.6B-GGUF google/gemma-4-E2B-it-qat-q4 0-gguf mlx-community/Qwen3.8-27B-4bit zai-org/GLM-5.3 来自我们博客的更多文章 Jun Kim，oMLX 创建者兼维护者，加入 Hugging Face 以支持 MLX 社区 +1 你本会自己提交的那个 PR 本文提到

匹配字段正文

码逻辑原封不动交了出来，直接替所有做私有化部署的团队抹平了本地环境和官方接口之间的行为温差。 这就像一家米其林餐厅，以前开源模型权重只是把菜谱免费印给你，你自己在家炒总是差一口锅气， 现在 DeepSeek 连自家后厨定制的高压猛火灶台 DeepInfer，外加主厨切配摆盘的标准化工序 deepseek-recipe 一块打包送你了。 不过我顺手去扒了一遍 GitHub 仓库的 README，发现对着全网欢呼的完全开源也得留个清醒，仓库文档写得很实在， 这次放出来的其实只是离线推理引擎，真正撑起完整线上高并发服务最核心的三块基建：KV 缓存系统、预填充与解码分离部署、多副本负载均衡这次并没有放出来；而且硬件门槛极高，目前只认英伟达最新的 Blackwell 架构 SM100 和华为昇腾 950 两款顶级算力卡，支持的模型也只限刚发的 V4.1-Flash，普通玩家手里的消费级显卡压根跑不起来。 但恰恰是 README 里这行只支持英伟达 SM100 和华为昇腾 950 的硬件限定，暴露了比框架本身更狠的产业信号，头部国产大模型已经不再等芯片厂商来缓慢适配自己，而是亲自下场写底层算子，一

匹配字段正文

低。 国内维度更有意思。这是主流大模型厂商里少见的"自研搜索引擎对外开放"动作——此前联网搜索要么是模型的内部能力不单独卖，要么走第三方数据源。DeepSeek等于宣布：我不只做模型，还自己养了一套搜索基础设施，并且愿意把它变成一门独立生意。 往深想一层 这笔账对DeepSeek来说是稳赚的。Agent调用搜索是刚性需求，尤其在编程场景里——查API文档、看GitHub issue、追最新库的用法，这些动作每天都在Claude Code和Codex里成规模发生。免费期相当于培养习惯，11月1日开始收费时，切换成本已经很高了。更妙的是，搜索数据本身会反哺DeepSeek对"用户真正在查什么"的理解，这在模型迭代里是稀缺信号。 但也要看到硬币的另一面。自研搜索引擎的索引规模、 freshness、反爬能力，和Google、Bing这种几十年的积累不在一个量级——DeepSeek没有公布任何搜索质量指标，目前所有判断都只能基于"能用"而非"好用"。每秒5次的限速对高并发Agent集群也是个硬约束，企业级负载大概率要排队。另外，把搜索和抓取绑定在自家API里，意味着开发者的数据管道进一步向单一

匹配字段正文

preparation step for FP4 inference. Getting started requires the latest Together SDK and the documentation. Together published cookbooks in a post-training-cookbook repository on GitHub covering GRPO training on the GSM8K benchmark, DPO implemented as a custom loss, and training multiple LoRA adapters within the same model resource; the documentation describes the cookbook as a Python library of ready-made loops whose notebooks walk through GRPO, SFT, and multi-adapter training on one model res

匹配字段正文

nt wrote and evolved it from under 10 to over 30 frames per second at 512x512 on WebGPU. See that the complete implementation and the web app are open source in the LiteRT Samples GitHub repo. What is the LiteRT Tensor API? The LiteRT Tensor API provides four key features for authoring and optimizing on-device models: Full control over graph structure: you decide every op in the model graph and every connection between them, from elementwise math to accelerator-optimized LLM composite ops, such

匹配字段正文

署应用和服务、制作病毒式社交媒体内容来推广产品、给供应商发邮件搭建电商渠道、致电企业为其搭建网站。只要还能负担运营成本，Automaton就会一直运作下去。 不过这款产品上线后评价两极分化。一方面它收获了大量关注。产品刚推出时，Sigil Wen就在X平台发文称，Conway Automaton注册智能体数量已经突破1.8万个；时至今日，Automaton的GitHub开源项目累计收获7300星。 另一方面质疑声同样不少。以太坊创始人Vitalik Buterin曾公开表态，认为所谓Web4的发展方向存在偏差。在他看来，拉长人类与AI之间的反馈链路，对整个世界而言并非好事。Softswiss首席信息官Denis Romanovskiy当时也提出，AI智能体能否实现真正的经济自主，取决于大模型在成熟度、记忆、规划以及工具调用上的能力；而在这些维度上，现有技术尚不稳定，不足以支撑可靠的无人值守运行。 结语：性能不及Muse、Dot，个人Agent跑在终端仍处早期探索阶段 这一轮个人智能体竞赛，主流思路仍是搭建云端虚拟机，让AI在厂商的服务器里持续执行任务。但Agent一旦开始高频循环、反复

匹配字段正文

rks, and CUDA-X libraries Local coding agents come to RTX Spark RTX Spark devices are ideal for local coding agents. NVIDIA has collaborated closely with the top agents, including GitHub Copilot, Pi, Hermes, and OpenClaw, to improve them and make them easier to use. The improvements include: Seamless deployment of local models with optimized recipes Enhanced security with NVIDIA OpenShell and/or Microsoft Execution Containers, delivering OS-level sandboxing Hybrid prompt routers between local an

匹配字段正文

spect the reflection-beam-jumpstart.yaml manifest template before configuring the deployment. Clone the sample repository, and enter the HyperPod Beam directory. git clone https://github.com/aws-samples/sagemaker-genai-hosting-examples.git cd sagemaker-genai-hosting-examples/SageMakerHyperpod/reflection-beam/ Create the local configuration files described in the README. cp config/deployment.env.example config/deployment.env cp config/release.env.example config/release.env Validate the JumpStart

匹配字段正文

isky software packages, extensions, and AI tool configurations. Repository and Cloud Availability SPACE, short for Sandboxed Platform for Agentic Code Execution, is published in a GitHub repository under the Apache License 2.0. The repository describes a platform that runs code in isolated Firecracker microVMs, with an HTTP API used to create Linux environments, execute commands, collect output, and manage sandbox lifecycles. A template supplies the starting environment, and a sandbox is an inst

匹配字段正文

hat record into skills, production-based evals, and fine-tuning datasets, all exportable along with the raw trajectories. CopilotKit is open source, with more than 38,000 stars on GitHub. TL;DR Agent tracing shows where an agent went wrong, while the correct path lives in what users do in the product, which most stacks never capture. CLiF, now in early access, learns from full-stack trajectories: the agent trace and the user’s actions in the product, securely captured as one timeline over AG-UI.

匹配字段正文

lugins/) a Dockerfile plus a runnability test. CI enforces the standards on every PR, and a suite of coding-assistant skills in the repo (prepare-python-recipe, generate-manifest, github-pr-review, and others) does most of the preparation, so your agent can scaffold, validate, and submit a recipe end-to-end. Built for coding assistants, not just developers Every AGENTS.md, every manifest.yaml, every SKILL.md is written primarily for the AI coding assistant that will open the repo on your behalf;

匹配字段正文

ng together again on the canvas. Whether you’re converting existing designs into working code, uploading from a local folder, bringing in a coded prototype via MCP, importing from GitHub, or capturing what’s live on your website with Figma’s Chrome extension, code layers lets you explore code-backed ideas on the canvas and push your thinking forward. And over time, we plan to make it possible to work with more complex codebases, too. Code layers is rolling out in open beta starting today, and is

匹配字段正文

An Apache-2.0 app with 4.8K GitHub stars is trying to make AI-assisted science as traceable as checksummed, versioned code. @aipoch ai’s Open-Science is a free, open-source desktop app where an AI agent reads papers, pulls data from science databases, writes and runs Python or R code, and saves the charts and reports, all in a single workspace on your computer. Research is usually split across PDFs, spreadsheets, code editors and terminals, so results lose their link to the data and code behind

匹配字段正文

する--serverも用意されている。 MCPツールはCode Mode経由の実行がデフォルトに Code Modeは、エージェントがJavaScriptを書き、利用可能なツールを組み合わせて実行する仕組み。1.xにも実験機能として存在したが、2.0では、外部ツールを提供するMCPサーバーのツールを呼び出す際にデフォルトで使う方式になった。 公式ブログでは、GitHubのバグ報告を取得し、Linearに課題として登録する例を示している。エージェントが一連の処理をJavaScriptで書き、OpenCodeのexecuteツールがそのコードを実行する。処理を一つのスクリプトにまとめることで、ツールを呼び出すたびにモデルへ結果を返し、次の呼び出しを判断させるやり取りを減らせるようになる。 別のツールの結果を待つ必要がない処理は、並列に実行できる。また、途中で得たデータをスクリプト内で加工し、必要な結果だけをモデルに返すこともできる。ブログは、多数のツールを接続した場合に、モデルへ渡すツールの説明などの情報量を抑えられる点も挙げている。 OpenCodeは、JavaScriptの実行に@op

匹配字段正文

g defensive security work, and static scanners miss the logic flaws that only surface once an application is running. Customers connect a repository through Chainguard’s Guardener GitHub app. From there, Harness compiles the application as it would ship, and the build boots in an isolated virtual machine cut off from the outside network. A frontier model then probes that running copy the way an attacker would. Nothing gets reported until it reproduces in a fresh environment on every replay, acco

匹配字段正文

ational element of Microsoft’s AI infrastructure strategy,” and said that Maia 200 chips are doing inference for Copilot, helping the Microsoft AI team to develop new models, with GitHub Copilot and Microsoft Foundry shifting to Maia 200s by the end of the year. (GitHub Copilot was supported by Maia 100s previously.) The Maia 200s will run OpenAI’s various models, Microsoft’s own models, and various open source models through Microsoft Foundry. It is not clear how much of OpenAI’s GPT API servic

匹配字段正文