149 个 · LLM 基础设施 · AI 核心
以孟加拉语优先、面向方言的 LLM 研究。开源孟加拉语 tokenizer，性能优于 Sarvam、AI4Bharat 和 GPT-4o（fertility 1.52，几乎零破损 conjuncts）。非商用，保护孟加拉语及其方言。Bengali-first, dialect-aware LLM research. Open-source Bengali tokenizer that outperforms Sarvam, AI4Bharat, and GPT-4o (fertility 1.52, near-zero broken conjuncts). Non-commercial, preserving Bengali and its dialects.
🚀 通过 70 个生成式 AI 实战项目转型技能，从入门到生产级架构师，涵盖真实场景应用。🚀 Transform your skills with 70 hands-on projects in Generative AI, guiding you from beginner to production-ready architect with real-world applications.
大语言模型元数据的开放注册表——以单一机器可读的 models.json 提供身份、作者、模态、上下文/输出限制、能力与生命周期日期，并通过 JSON Schema 校验。CC BY 4.0Open registry of large-language-model metadata — identity, authorship, modalities, context/output limits, capabilities & lifecycle dates as one machine-readable models.json validated by JSON Schema. CC BY 4.0.
🛠️ 通过 Local-LLM（一款基于 BERT 模型的轻量级 Python 库）在安全环境中实现离线 NLP 工作流，确保可复现性与可靠性。🛠️ Enable offline NLP workflows with Local-LLM, a lightweight Python library for secure environments using the BERT model, ensuring reproducibility and reliability.
基于 TurboQuant 优化 FAISS 兼容的向量量化，实现快速、精准的向量检索。Optimize FAISS-compatible vector quantization for fast, accurate vector search with TurboQuant
轻松优化面向 AI 系统的 prompt，借助直观工具与特性提升性能🚀 Optimize your prompts for AI systems easily and boost performance with intuitive tools and features designed for better results.
《大模型推理原理与优化》：面向系统/架构/后端研发工程师的模型原理入门课，目标是通俗易懂的解释推理过程，理解原理有助于系统开发/维护工作
DP-ES 官方实现：面向 Prompt 优化的差分隐私进化策略（EMNLP 2026）。Official implementation of DP-ES: Differentially Private Evolution Strategies for Prompt Optimization (EMNLP 2026)
混合神经符号 AI：Llama 3.2 1B + 精确数学 + 经验证事实，807 MB 即时 CPU 原生回答，无需 GPU。封装 .aef 分发。Hybrid neuro-symbolic AI: Llama 3.2 1B + exact math + verified facts — instant CPU-native answers in 807 MB, no GPU. Sealed .aef distribution
通过一次 wrap() 调用为 LLM 流水线构建运行时可靠性守卫，可配置地防御常见生产故障Build runtime reliability guards for LLM pipelines with one wrap() call and configurable protection against common production failures
WordPress 共享 AI 基础设施：provider 凭证、实时模型、提示词、标准化请求及 AI-Scribe 和兼容插件的使用记录。Shared AI infrastructure for WordPress: provider credentials, live models, prompts, normalised requests and usage records for AI-Scribe and compatible plugins.
论文 A Survey of On-Policy Distillation for Large Language Models（arXiv:2604.00626）的配套网站。Companion website for A Survey of On-Policy Distillation for Large Language Models (arXiv:2604.00626).
使用纯 C 从零构建 LLM 推理引擎，不依赖任何框架。Build an LLM inference engine from scratch in pure C with no frameworks.
论文《Controllable molecular graph generation from natural-language chemical constraints》（基于自然语言化学约束的可控分子图生成）的代码仓库。This is repository for "Controllable molecular graph generation from natural-language chemical constraints"
几何有限主义：语言与思维的几何学Geofinitism: The Geometry of Language and Thought
构建 LongCat-Flash-Prover，利用 LongCat 模型进行快速定理证明与形式化推理。Build LongCat-Flash-Prover for fast theorem proving and formal reasoning with LongCat models
使用纯 C99 MoE 推理引擎在 CPU 上原生运行 DeepSeek-V4-Flash-0731，无需 GPU、CUDA 或 PyTorch。Run native DeepSeek-V4-Flash-0731 on CPU with a pure C99 MoE inference engine — no GPU, CUDA, or PyTorch needed.
构建安全、有治理、可观测、成本可控的云与 AI 平台能力的实用参考框架。A practical reference framework for building secure, governed, observable, and cost-aware Cloud & AI platform capabilities.
The Convergence Gap 的代码与 artifact：指令微调模型何时收敛到下一 token 预测Code and artifacts for The Convergence Gap: when instruction-tuned models settle on next-token predictions.
生产级 LLM 网关：兼容 OpenAI 的 API，支持语义缓存，可在 Claude/OpenAI/Ollama 之间进行智能路由与成本分析。Production LLM gateway: OpenAI-compatible API, semantic caching, intelligent routing across Claude/OpenAI/Ollama, cost analytics.
使用 TurboQuant 和 MLX 在 Windows 上运行 Qwen3.5 语言模型，实现快速的本地推理。Run the Qwen3.5 language model on Windows using TurboQuant and MLX for fast local performance.
Experimental Qwen3.5-derived 752M LLM：基于 Qwen3.5 的实验性 752M 参数 LLM，采用 CPT + SFT 两阶段流程，面向编码、技术推理与指令跟随。Experimental Qwen3.5-derived 752M LLM: a two-phase CPT + SFT pipeline for coding, technical reasoning, and instruction following.
通过 MegaQwen CUDA megakernel 加速 Qwen3-0.6B 推理，在 RTX 3090 上达到 531 tok/s decode，较 HuggingFace 提升 3.9×🚀 Achieve faster Qwen3-0.6B inference with the MegaQwen CUDA megakernel, delivering 531 tok/s decode on RTX 3090—3.9x faster than HuggingFace.
大语言模型（LLM）是一种人工智能（AI）程序，能够识别并生成文本，以及执行其他任务。A large language model (LLM) is a type of artificial intelligence (AI) program that can recognize and generate text, among other tasks.
在双 RTX 3090 GPU 上以 262K 上下文长度并发流运行 Qwen3.6-27B，开发工作已迁移至 club-3090。Run Qwen3.6-27B at 262K context with concurrent streams on dual RTX 3090 GPUs. Development moved to club-3090.
完全在浏览器中通过 WebAssembly 实现文档转换、检查、OCR 与翻译。保留版式的翻译，全程本地、离线优先。Convert, inspect, OCR, and translate any document entirely in the browser via WebAssembly. Layout-preserving translation, fully local, offline-first.
使用支持 macOS 和 Windows 平台的开源工具高效构建与定制机器学习模型Build and customize machine learning models efficiently with an open-source tool that supports macOS and Windows platforms.
MAI-Code 模型的官方仓库，用于发布 MAI-Code 模型版本与更新，并通过 issue 与反馈与开发者社区互动。Official repo for MAI-Code models, where we publish MAI-Code model releases and updates and engage with developer community on issues and feedback.
通过单一严格的 YAML 清单统一管理本地 LLM 运行时——支持状态查看、健康检查、启动、停止，基于原生 systemd 与 Docker。Manage local LLM runtimes from one strict YAML manifest — status, doctor, start, stop over native systemd and Docker.
metaScreener——基于插件的桌面应用，用于 human-in-the-loop systematic literature screening。在顺序可审计的 pipeline 中结合确定性启发式过滤与 LLM 推理，通过 SHA-256 校验包实现完全可复现。MIT 协议。metaScreener — a plugin-based desktop application for human-in-the-loop systematic literature screening. Combines deterministic heuristic filters with LLM inference in a sequential, auditable pipeline. SHA-256 verified bundles for full reproducibility. MIT licensed.
面向 LLM 的阶段感知上下文窗口治理框架。提供不变的上下文长度上限、基于熵的稳定性控制，以及针对降级（碎片化）状态的概率性保证，适用于生产级 LLM 系统。Phase-aware context window governance framework for Large Language Models (LLMs). Provides invariant context length caps, entropy-based stability control, and probabilistic guarantees against degraded (fragmentation) states for production LLM systems.
🛠 通过此硬件插件提升 vLLM 在 Kunlun XPU 上的性能，无缝集成主流 AI 模型并优化执行效率🛠 Enhance vLLM performance on Kunlun XPU with this hardware plugin, offering seamless integration for popular AI models and optimized execution.
适用于 DeepSeek、Qwen、GLM 及多种 AI 模型的 OpenAI 兼容 API 示例。OpenAI Compatible API examples for DeepSeek, Qwen, GLM and multiple AI models.
🚀 使用强化学习优化半精度通用矩阵乘法（HGEMM）CUDA kernel，性能超越 cuBLAS 及其他基准。🚀 Optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels using reinforcement learning, surpassing cuBLAS and other benchmarks with superior performance.