←搜索结果

搜索“Hugging Face”

按匹配度 · 共 1193 条结果 清除搜索

星期一

  1. 蚂蚁百灵：Developer Blog

    蚂蚁百灵发布并开源 Ling-2.5-1T 即时模型

    蚂蚁百灵发布并开源旗舰级即时模型 Ling-2.5-1T，总参数 1T、激活 63B，预训练语料从前代 20T 扩展至 29T，支持最高 1M token 上下文。模型采用 1:7 的 MLA + Lightning Linear 混合线性注意力架构，引入正确性与过程冗余复合奖励机制，在 AIME 2026 上平均输出约 5890 token 即接近消耗 15k–23K token 的前沿思考模型水平。官方称其在复杂推理与指令遵循上优于 DeepSeek V3.2、Kimi K2.5、GPT 5.2，可适配 Claude Code、OpenCode 等智能体产品，但在复杂智能体交互和长程任务上仍落后于前沿模型。

    rompt：请你写一个 minialipay，前后端分离架构 + 内存数据库存储，支持基础的用户登录、转账等功能，并且有新春特色的集五福功能，欢庆春节主题。 Opencode 接入（skills） 能够调用正确的 Skills，完整执行内部系统开发指令，Agent 自动化完成从基础架构搭建到多品牌主题逻辑的代码实现，将需求描述直接转化为具备导航、搜索及动态换肤功能的可运行原型。 Minecraft 机器人任务执行 精准执行空间构建指令，Agent 自动化完成从底层铺设到顶层收缩的数百次方块放置，将自然语言直接转化为游戏内的实体建筑。（使用的 Agent 框架 6. 局限性与未来计划 Ling-2.5-1T 已实现高吞吐解码与领先的超长上下文处理能力，并初步具备智能体交互能力，为通用智能体时代奠定基础。但在复杂智能体交互和长程任务上，它仍落后于前沿模型。下一版本将聚焦真实世界复杂任务的长程执行与交付效率优化，同时持续提升 Token 效率，以在效率与效果间取得更优平衡。 欢迎通过以下方式下载、访问和使用 Ling-2.5-1T。 🤗 Hugging Face： 🤖 ModelScope：

    匹配字段正文

    #engineering#模型发布#开源生态#推理
  2. Qwen Blog精选

    Qwen3.5 发布：首个原生多模态智能体模型 Qwen3.5-397B-A17B

    Qwen 团队正式发布 Qwen3.5 系列首个开源权重模型 Qwen3.5-397B-A17B，这是一个原生视觉语言模型，在推理、编码、智能体和多模态理解等基准上给出完整评测结果。模型采用线性注意力（Gated Delta Networks）与稀疏 MoE 混合架构，总参数 397B、每次前向仅激活 17B，语言与方言支持从 119 种扩展到 201 种。托管版本 Qwen3.5-Plus 通过阿里云百炼提供，默认 1M 上下文窗口，并内置官方工具与自适应工具调用。

    s | Qwen Blog Publication About Try Qwen Chat Qwen3.5: Towards Native Multimodal Agents February 14, 2026 · 196 min · 41706 words · Qwen Team | Translations: 简体中文 QWEN CHAT GitHub Hugging Face ModelScope DISCORD We are delighted to announce the official release of Qwen3.5 , introducing the open-weight of the first model in the Qwen3.5 series, namely Qwen3.5-397B-A17B . As a native vision-language model, Qwen3.5-397B-A17B demonstrates outstanding results across a full range of benchmark evaluatio

    匹配字段正文

    #company#模型发布#多模态#推理

星期二

  1. Prime Intellect精选

    Prime Intellect 发布 Lab 全栈训练平台

    Prime Intellect 发布 Lab，把 Environments Hub、Hosted Training 和 Hosted Evaluations 整合为面向智能体后训练的全栈平台，覆盖大规模智能体 RL、推理与评测。平台以环境为核心，包含任务数据集、模型 harness 和评分 rubric，当前 Hosted Training 支持基于 prime-rl 的 LoRA 智能体 RL，后续将加入 SFT、DPO 等算法。私有测试期间已完成 3000 多次 RL 运行，Environments Hub 累计有 250 多位创作者构建 1000 多个环境、下载量超过 10 万次。

    Introducing Lab: The Full-Stack Platform for Training your Own Models

    f currently available models for Hosted Training, do: Beyond our own INTELLECT-3 model, many awesome small and large models are available from organizations such as NVIDIA, Arcee, Hugging Face, Allen AI, Z.AI, Alibaba Qwen, and many more launching soon. We've also added experimental multimodality support and have deployed our first models to train on environments with image input. Runs are configured via a simple .toml file: To start a training, run: You can then view logs, training results, and

    匹配字段正文

    #company#产品更新#Agent#数据/训练

星期四

  1. Mistral AI精选

    Mistral AI 发布 Voxtral Transcribe 2，实时转写延迟低于 200ms

    Mistral AI 发布 Voxtral Transcribe 2，包括批量转写模型 Voxtral Mini Transcribe V2 和实时模型 Voxtral Realtime。两者支持 13 种语言；前者提供说话人分离、上下文偏置和词级时间戳，单次可处理最长 3 小时音频，API 价格为 $0.003/min。4B 参数的 Realtime 延迟最低可配置至 200ms 以下，API 价格为 $0.006/min，并以 Apache 2.0 许可开放权重。

    Fleur Voxtral 2

    h a 4B parameter footprint, it runs efficiently on edge devices, ensuring privacy and security for sensitive deployments. We’re releasing the model weights under Apache 2.0 on the Hugging Face Hub. Voxtral Mini Transcribe V2. Average diarization error rate (lower is better) across five English benchmarks (Switchboard, CallHome, AMI-IHM, AMI-SDM, SBCSAE) and the TalkBank multilingual benchmark (German, Spanish, English, Chinese, Japanese). Average word error rate (lower is better) across the top-

    匹配字段正文

    #lab#模型发布#语音#端侧

星期二

  1. Sarvam AI Blog

    Sarvam AI 发布 Sarvam Audio：基于 Sarvam 3B 的语音识别模型

    Sarvam AI 发布 Sarvam Audio，这是基于 30 亿参数语言模型 Sarvam 3B 的音频扩展，支持 22 种印度语言，提供五种可控转写格式（逐字转写、规范化非混码、规范化混码、罗马化输出、智能翻译）。在 IndicVoices 评测集上，Sarvam Audio 在非规范化、规范化和混码三种转写风格下的 WER 均优于 GPT-4o-Transcribe 和 Gemini-3-Flash，并支持最长 60 分钟音频的多说话人分离识别。

    Sarvam Audio: Speech Recognition beyond Transcription

    ates a consistent performance advantage across languages. The evaluation framework is open-sourced here, and the Synthetic Contextual ASR Benchmark (Indic) is publicly released on Hugging Face. Benchmark overview.This benchmark evaluates context-aware ASR for voice-bot interactions across the top 10 Indian languages. Each sample represents a single dialog turn and includes audio, ground-truth transcription, language, and rich conversational context (bot persona, history, and prompt). The dataset

    匹配字段正文

    #company#模型发布#语音#多模态

星期一

  1. LMSYS Blog

    SGLang RL 团队实现 INT4 QAT 端到端 RL 训练，单张 H200 跑约 1TB 模型

    SGLang RL 团队在 slime 框架上复现并落地了 INT4 QAT 端到端方案，训练侧用假量化加 STE，推理侧用 W4A16 真实量化，使约 1TB 规模的 K2 类模型可装入单张 H200（141GB）。消融实验显示，训练端假量化与推理端真实量化必须同时开启，否则训练与推理的 logprob 差异会明显增大；在 Qwen3-235B-A22B 和 Kimi-K2-Thinking 上，BF16 训练加 INT4 推理的奖励与评测分数走势与 BF16、FP8 方案基本重合。团队称 INT4 的主要收益来自显存压缩带来的单节点 rollout，因为 H 系列 GPU 没有原生 INT4 Tensor Core，其计算速度与 FP8 大致同档。

    End-to-end QAT INT4 pipeline

    its built-in Marlin kernel as the INT4 inference backend. However, in practice we encountered a notable “format gap”: QAT training outputs weights in standard formats (similar to Hugging Face), while SGLang’s Marlin kernel requires weights to be specially packed and permuted so that the kernel can read them efficiently. Given that RL training requires frequent weight updates, we must solve format compatibility. We therefore designed a reverse restore weights before loading safety mechanism. Usi

    匹配字段正文

    #company#开源/仓库#数据/训练#部署/工程
  2. Qwen Blog精选

    Qwen 发布旗舰推理模型 Qwen3-Max-Thinking

    Qwen 团队发布旗舰推理模型 Qwen3-Max-Thinking，通过扩大参数规模并投入大量算力做强化学习，在事实知识、复杂推理、指令遵循、人类偏好对齐和智能体能力上均有提升，在 19 项基准上表现与 GPT-5.2-Thinking、Claude-Opus-4.5、Gemini 3 Pro 相当。模型新增自适应工具调用能力，可在对话中自主调用内置的 Search、Memory 和 Code Interpreter，已在 chat.qwen.ai 上线；同时采用经验累积的多轮测试时扩展策略，在 GPQA、HLE、LiveCodeBench v6 等基准上优于标准并行采样。API 模型名为 qwen3-max-2026-01-23，兼容 OpenAI 与 Anthropic 接口协议，可直接配合 Claude Code 使用。

    Bench 38.2 56.3 51.6 44.1 40.9 Deep Planning 6 44.6 33.9 23.3 21.6 28.7 Long Context AA-LCR 72.7 74.0 70.7 65.0 68.7 We evaluated only on the text subset. ↩︎ We blocked access to Hugging Face and other HLE-related websites to prevent data leakage. ↩︎ For reproducibility of Arena-Hard v2, we report the win rates evaluated by GPT-4.1. ↩︎ We followed the official setting of Tau² Bench with no custom scaffolding. ↩︎ BFCL-V4 is configured with a maximum of 100 interaction turns. ↩︎ Deep Planning is

    匹配字段正文

    #company#模型发布#推理#Agent

星期三

  1. LMSYS Blog

    Novita AI 优化 GLM4-MoE 生产部署：TTFT 降低 65%，TPOT 提升 22%

    Novita AI 基于 SGLang 为 GLM4-MoE 模型开发了一套生产级优化方案，在 H200 集群 TP8 与 FP8 配置下实现 TTFT 最高降低 65%、TPOT 提升 22%。优化包括 Shared Experts Fusion、Qknorm Fusion 和 Async Transfer，其中 Shared Experts Fusion 单独带来 TTFT 最高 23.7%、ITL 20.8% 的提升。Suffix Decoding 在 MTP 加速基础上进一步将 TPOT 从 25.13ms 降至 19.63ms，相关组件多数已合并至 SGLang 上游。

    Highly structured agentic behaviors: Fixed phrases like "Let me...", "Now let me..." appear frequently To support further research, we have open-sourced the evaluation dataset on Hugging Face : Agentic Code Dataset on Hugging Face Performance Comparison With built-in MTP acceleration, Suffix Decoding further reduces TPOT by 22% (from 25.13ms to 19.63ms): Conclusion The combination of these optimizations provides comprehensive performance improvements for SGLANG deployments: Most components are

    匹配字段正文

    #company#教程/实践#部署/工程#开源生态

星期五

  1. Qwen Blog

    Qwen3-TTS 更新：49 种音色、10 种语言与 9 种方言

    Qwen 团队发布 Qwen3-TTS-Flash 更新，支持 49 种以上音色、10 种语言和 9 种方言的语音合成，可通过 Qwen API 调用。官方称在 MiniMax TTS 多语言测试集上，其平均词错误率低于 MiniMax、ElevenLabs 和 GPT-4o-Audio-Preview。相比上一版本，模型根据文本自适应调整语速和韵律的能力明显提升，更接近真人语音。

    Qwen Blog Publication About Try Qwen Chat Qwen3-TTS Update! 49 Timbres + 10 Languages + 9 Dialects December 5, 2025 · 6 min · 1086 words · Qwen Team | Translations: 简体中文 DASHSCOPE HUGGING FACE DEMO MODELSCOPE DEMO Qwen3-TTS-Flash is a flagship text-to-speech model that supports multi-timbre, multi-lingual, and multi-dialect speech synthesis. It aims to produce natural and expressive speech and is available via Qwen API . Major Improvements: Richer Timbres Support : Qwen3-TTS offers over 49 high-

    匹配字段正文

    #company#模型发布#语音#多模态

星期五

  1. vLLM 官方博客

    NVIDIA Nemotron Nano 2 VL 发布并获 vLLM 支持

    NVIDIA Nemotron Nano 2 VL 开源视觉语言模型发布，面向视频理解与文档智能，并获 vLLM 推理支持。模型采用混合 Transformer–Mamba 架构，通过 EVS 剪除冗余视频帧以减少 token。vLLM 夜间版本支持以 BF16、FP8 和 FP4 精度部署 OpenAI 兼容的推理服务器。

    r; Efficient video sampling as token compression module; Hybrid Transformer-Mamba Architecture - Nemotron Nano 2 LLM backbone with reasoning; Accuracy: Download model weights from Hugging Face - BF16, FP8, FP4-QAD; Run with vLLM for inference; Technical report to build custom, optimized models with Nemotron techniques Run optimized inference with vLLM This guide demonstrates how to run Nemotron Nano 2 VL on vLLM, achieving accelerated inference and serving concurrent requests efficiently with BF

    匹配字段正文

    #engineering#模型发布#多模态#推理

星期二

  1. vLLM 官方博客精选

    vLLM 上 Kimi K2 工具调用准确率的深度调试

    vLLM 与 Kimi 团队修复了 Kimi K2 的关键工具调用兼容问题，使成功解析的调用数从 218 个增至 971 个，提升 4.4x。问题涉及缺失 add_generation_prompt、空 content 的结构转换，以及历史工具调用 ID 不符合 functions.func_name:idx 格式。修复后仍出现 316 次 schema_validation_error_count，团队正协作将用于约束工具选择的 Enforcer 组件集成到 vLLM。

    x: After identifying the root cause, I collaborated with the Kimi team. They were incredibly responsive and, based on my findings, updated the model's tokenizer config.json on the Hugging Face Hub. The fix was to explicitly declare add generation prompt as a supported parameter for the chat template. This allowed vLLM to pass the argument correctly, fixing the primary source of failed tool calls. Additionally, I submitted this PR, where I whitelist standard chat-template parameters when tokenize

    匹配字段正文

    #engineering#教程/实践#MCP/工具调用#部署/工程

星期一

  1. vLLM 官方博客

    vLLM Semantic Router 如何通过可扩展 LoRA 从单体架构走向模块化

    vLLM 重构 Semantic Router 的 Rust 分类层，以模块化架构、共享基础模型的 LoRA 适配器和并发优化降低多任务分类开销。Qwen3-Embedding 支持 100 多种语言和 32,768 tokens 上下文；OnceLock 实现初始化后无锁读取，Flash Attention 2 可为 Qwen3 注意力操作带来最高 4× 加速。

    ements. Long-Context Embedding Models Two new embedding models address the context length limitation: Qwen3-Embedding Qwen3-Embedding supports context lengths up to 32,768 tokens (Hugging Face model card). The implementation uses a RoPE (Rotary Position Embedding), enabling this extended context handling through improved frequency resolution at longer distances. Qwen3-Embedding was trained on text from over 100 languages (Hugging Face model card), making it suitable for multilingual routing scen

    匹配字段正文

    #engineering#产品更新#部署/工程#MCP/工具调用

星期三

  1. Preferred Networks：AI 研究与产品

    Preferred Networks 如何改进 MinHashLSH 实现大规模数据集去重

    Preferred Networks 预训练团队实习生黄宝玺撰文分析 MinHashLSH 在大规模 LLM 训练数据去重中的局限：候选对匹配难以并行化，且需为所有文档存储 MinHash 签名，5 亿文档规模约需 23 TB 空间。文章指出，精心去重后模型可用少 13–41% 的 token 达到相近或更好的准确率。

    o address this, we turned to a larger public dataset as shown in Figure 6 and constructed subsets ranging from 10 GB to 40 GB by duplicating documents. Figure 6. Public dataset in Hugging Face [11] 5.1 Precision, Recall and F1 score When evaluating a deduplication method, three key metrics are considered: precision, recall, and computational performance. Precision measures how many of the predicted duplicates are actually correct. Low precision indicates excessive false positives, which is harmf

    匹配字段正文

    #company#教程/实践#数据/训练#部署/工程

星期四

  1. vLLM 官方博客

    vLLM Semantic Router 推进 LLM 智能推理路由

    vLLM Semantic Router 为 vLLM 推理引擎提供开源的意图感知路由层，按查询复杂度选择快速路径或 CoT 推理模式。试验中，该设计使准确率提高约 10%、延迟降低约 50%、token 减少约 50%，并支持通过 Kubernetes 和 Envoy 部署。v0.1 路线图包括语义缓存、多因素路由、性能测试和路由策略可视化。

    determine routing paths; Smart Routing: Simple queries → "fast path" inference; Complex queries → "Chain-of-Thought" reasoning mode; High-Performance Engine: Written in Rust using Hugging Face Candle, it delivers high concurrency and zero-copy inference; Cloud-Native Integration: Works out-of-the-box with Kubernetes and Envoy via the ext proc plugin In trials, this design yielded: 10% higher accuracy; 50% lower latency; 50% fewer tokens In business and economics domains, gains exceeded 20% accur

    匹配字段正文

    #engineering#产品更新#推理#部署/工程

星期一

  1. Cato AI Labs

    Aim Labs 推出新型动态模型扫描器，弥补静态扫描防护模型供应链攻击的固有缺陷

    Aim Labs 提出一种新的专利待审模型扫描方法，通过在沙箱环境中执行模型并追踪其操作，弥补现有静态扫描器的固有缺陷。现有扫描器依赖预设函数与模块黑名单，难以覆盖数百万 Python 函数，且静态代码分析属 NP-hard 问题，容易被绕过。该方法已在 Black Hat USA 2025 演讲“Smashing Model Scanners: Advanced Bypass Methods and a Novel Detection Approach”中展示。

    Executive Summary The AI transformation relies on usage of AI models that implement complex algorithmic tasks. Repositories such as Hugging Face are home to millions of pretrained models. Anyone can upload any model, and anyone can download. Using untrusted third-party models introduces a supply chain risk of code execution, either in model load time or in inference time – making AI security essential to ensure model integrity and protect against these emerging risks. To mitigate those risks, Hu

    匹配字段正文

    #research#安全/对齐#开源生态#Hugging Face

星期二

  1. vLLM 官方博客

    vLLM 新增对 gpt-oss 的支持

    vLLM 现已支持在 NVIDIA Blackwell、Hopper 以及 AMD MI300x、MI355x GPU 上运行 gpt-oss。其针对 MXFP4 MoE、高效注意力和混合 KV cache 集成了专用内核与内存管理方案，并通过 OpenAI Responses API、gpt-oss toolkit 或外部 MCP 工具服务器支持工具调用。

    . The AMD team contributed significantly to the integration of the model on their devices: Hongxia Yang, Ali Zaidy, with great support from Peng Sun, Vinayak Gokhale, Andy Luo The Hugging Face team continues to be amazing at building an open source ecosystem: Lysandre, Hugo, Marc, vb, Arthur, Mohamed, Andrien. Finally, we want to thank all the partners that leveraged vLLM in some way and delivered valuable feedback and improvements to this effort: AWS, Cloudflare, Snowflake, Databricks, Together

    匹配字段正文

    #engineering#产品更新#部署/工程#MCP/工具调用

星期一

  1. vLLM 官方博客

    MiniMax-M1 混合架构如何借助 vLLM 实现长上下文快速推理

    vLLM 通过内存管理与底层计算优化支持 MiniMax-M1 的混合架构，实现长上下文高效推理与部署。MiniMax-M1 总参数量为 456B，推理时动态激活约 45.9B 参数；其 Lightning Attention 已通过 Triton 集成至 vLLM。文章还提供了使用 vLLM 和 Docker 部署该模型的快速指南。

    throughput; Efficient and intelligent memory management; Robust support for batched requests; Deeply optimized backend performance Model Download You can download the models from Hugging Face: Deployment Below is a quick guide to deploying MiniMax-M1 with vLLM and Docker: MiniMax-M1 Hybrid Architecture Highlights Mixture-of-Experts (MoE) MiniMax-M1 utilizes a Mixture-of-Experts (MoE) architecture with 456 billion total parameters. During inference, a dynamic routing algorithm activates a sparse

    匹配字段正文

    #engineering#教程/实践#部署/工程#推理

星期二

  1. Liquid AI 模型与工程博客

    LFM-1B-Math：小模型能否成为简洁推理者？

    Liquid AI 基于 1.3B 的 LFM 模型，通过约 4.5M 样本的监督微调和短程强化学习，训练出 LFM-1.3B-Math，使其在有限响应预算下具备简洁数学推理能力。该模型无需数学专用预训练，适合内存和生成速度受限的边缘设备部署。

    dditional 3M samples. We found Nvidia’s OpenMathReasoning [5] and Llama Nemotron Post-Training [6] datasets particularly effective among other openly available datasets, including Hugging Face’s Open-R1 [7] and Bespoke Lab’s OpenThoughts (e.g., OpenThoughts2-1M) [8]. We selected OpenMathReasoning, which we found to yield strong empirical results on math problems, and combined it with the science subset of Llama Nemotron Post-Training and SCP-116K [9] to enable applicability to science problems.

    匹配字段正文

    #company#论文/研究#推理#数据/训练

星期五

  1. Sarvam AI Blog

    Sarvam AI 发布 Sarvam-M：基于 Mistral Small 的 24B 混合推理 Indic 语言模型

    Sarvam AI 基于 24B 的 Mistral Small（Apache 2.0 许可）经 SFT 与 RLVR 后训练，推出混合推理模型 Sarvam-M，在印度语言基准上平均提升 20%、数学基准提升 21.6%、编程基准提升 17.6%，罗马化印度语言 GSM-8K 提升 86%。该模型在多数基准上超过 Llama-4 Scout，与 Llama-3.3 70B、Gemma 3 27B 相当，但 MMLU 等英文知识基准下降约 1 个百分点。团队还用 TensorRT-LLM 完成 FP8 量化与 lookahead decoding，并在 H100 上部署于 NVCF。

    Home; /; Blog Explorations in Post Training and Inferencing Optimizations for a Hybrid Indic LLM Download the model from Hugging Face, try it on our playground, and build with our APIs. Towards Building a Sovereign AI Ecosystem in India, we plan to have regular model drops and share our detailed technical findings. This is the first in this series of technical blogs, wherein we share our findings on post-training. We look forward to hearing feedback and suggestions for collaborations. In this bl

    匹配字段正文

    #company#模型发布#开源生态#数据/训练

星期日

  1. Prime Intellect精选

    Prime Intellect 发布 INTELLECT-2：首个通过全球分布式强化学习训练的 32B 模型

    Prime Intellect 发布 INTELLECT-2，称其为首个通过全球分布式强化学习训练的 32B 参数模型，在动态异构的算力贡献者集群上以完全异步 RL 训练推理语言模型。团队为此从零构建了异步 RL 框架 PRIME-RL、验证推理 worker rollout 的 TOPLOC 和广播策略权重的 SHARDCAST，并修改标准 GRPO 训练配方与数据过滤方法以稳定训练，在数学和编码基准上提升了 QwQ-32B 的表现。模型、代码和数据均已开源，技术报告见 primeintellect.ai/intellect-2。

    INTELLECT-2 Release: The First 32B Parameter Model Trained Through Globally Distributed Reinforcement Learning

    with our code and data, hoping to enable more open research in the field of globally-distributed training Detailed Technical Report: primeintellect.ai/intellect-2; INTELLECT-2 on Hugging Face Chat Interface to try it out: chat.primeintellect.ai; prime-rl: – our async RL framework Paradigm Shift for Distributed Training Test-time compute scaling with reinforcement learning has emerged as a new scaling axis for large language models (LLMs), enabling improvements by allowing models to spend more t

    匹配字段正文

    #company#模型发布#数据/训练#开源生态