LLM & Inference

零基础入门大模型（LLM）的基础教程：从零手写 GPT-2，一行行看懂大模型是怎么工作的 | Build GPT-2 from scratch in PyTorch, step by step

Local-first engineering agent that deploys, benchmarks and tunes vLLM/SGLang inference services on your own GPUs, with auditable run evidence.

A complete collection of RAG interview questions, answers (1619 questions & 52 RAG types), system design scenarios, architecture patterns, and production-ready concepts.

不想啃 5000+ 全文？我已经替你和 LLM 啃完了 — ICLR 2026 全景中文导读

LLM agent 的有界自纠错长期记忆层：把连续、含矛盾的项目交互日志蒸馏成有界、可自更新的记忆库。Bounded, self-correcting long-term memory for LLM agents — candidate pool + value dynamics + tension adjudication + causal-replay evals.

Turns vibe coding into high quality code. Sets up rules and gates in your repo so AI-written code stays correct without you reviewing every line.

Hand-written NVFP4 W4A16 CUDA kernels for Volta

精酿 BrewReel：让 DeepSeek 这类便宜模型也能做出好看的竖版宣传片。写一份产品简报，AI 挑镜头、写文案，一条命令出片。3 种配方、6 个行业、广告法校验，开源可商用。

Command a fleet of AI agents, on every deck, every desk, every wrist. Self-hosted multi-agent platform with memory, a coding pipeline, Flows, a team kiosk and voice devices. Any LLM, or 100% local with Ollama.

M-Courtyard: Local AI Model Fine-tuning Assistant for Apple Silicon. Zero-code, zero-cloud, privacy-first desktop app powered by Tauri + React + mlx-lm.

Stop your coding agent from stalling real work on self-invented bookkeeping - receipts, hashes, locks, certification rituals. Ship first, then verify. Skill for Claude Code, Codex, and other agents.

DSH plugin — free OpenCode Zen models for DeepSeek Harness (DSH). Free LLM API, no API key needed. 在 DSH 中使用 OpenCode Zen 免费模型，无需 API key

Production-grade recipe for DeepSeek-V4-Flash-0731 (284B MoE) on 2x NVIDIA DGX Spark: self-healing 2-node vLLM cluster, reboot-verified, tuned DSpark speculative decoding (~75 tok/s), full benchmarks, OpenAI Codex CLI integration

👁️ Review AI generated code and plan changes

🐋 Полный гайд на русском: как установить и запустить DeepSeek локально (V4, R1) на ПК, Mac и сервере — Ollama, llama.cpp, LM Studio, vLLM, Open WebUI, API, RAG, агенты, fine-tuning, лучшие фишки

辞达：用你选择的大模型，在 Mac 的任何地方翻译和润色文字。Translate and polish text anywhere on your Mac.

Community registry of LLM serving-path traps that produce confidently wrong measurements: templates, tool parsers, reasoning fields, quant kernel paths, CUDA toolchains, KV allocation, eval harnesses, versioning. Symptom-first, with the check that catches each.

LLM infrastructure cost reduction via NUMA-aware weight banking: 147 t/s (8.8x stock llama.cpp) on refurbished enterprise POWER8. Self-hosted inference, no cloud APIs. Part of the Proof of Physical AI stack.

Feel how fast a local LLM runs before you buy the hardware

A Large-Scale Knowledge Graph for Automated Scientific Research

Survey companion for Computational Character Modeling: role-playing methods, data foundations, and evaluation, organized by the manuscript.

Dyno Lab — an AI safety, alignment and interpretability research workbench for Apple Silicon. Explore activations, probes, SAEs and interventions with local inference, a Python SDK, APIs and MCP.

Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses | 500+ Papers | Perception, Cognition, Planning, Interaction, Agentic System

An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682

Multi-model council for DeepSeek Harness: independent answers, anonymous peer reviews, and an inspectable final decision.

[ICML 2026] d3LLM: Ultra-Fast Diffusion LLM 🚀

The fastest way to run Qwen3.8 27B on Strix Halo (gfx1151)

🤗 [TMM 2026] Official implementation for "CC-Pan: Channel-wise Compression based Diffusion for Efficient Pan-Sharpening" https://arxiv.org/abs/2602.04473

Autonomous AI agent that plays Minecraft on its own — a server-side Fabric mod (MC 1.21.3) where a DeepSeek / OpenAI-compatible LLM drives a real player to mine, build, farm, fight and survive from natural-language commands. LLM plans, deterministic tasks execute — no Python harness, no bot account.

GLM-5.2-NVFP4-REAP-469B serving on SM120 (4× RTX PRO 6000 Blackwell) — one-command vLLM launch recipe, 250K context, DeepSeek Sparse Attention + MTP speculative decode

PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers

[NeurIPS 2026] State of Thought Enables Endogenous Reasoning

DeepSeek-V4-Flash-0731 284B inference in ~25 GB of RAM / Qwen3.8-Next-Flash inference in ~15 GB of RAM on any M-series MacBook

Upload your data → Get a fine-tuned SLM. Free.

Heterogeneous prefill/decode for DeepSeek-V4-Flash: CUDA prefill (DGX Spark, vLLM) -> Metal decode (Mac Studio, oMLX) over plain 10GbE

A model-agnostic GPU runtime for shipping models as verified programs.

Speedrunning LoRA fine-tuning: frozen task, frozen hardware, public wall-clock leaderboard. modded-nanogpt for fine-tuning.

Unofficial skill that teaches coding agents to build with TypeSafe AI's Jev: typed decisions, calibrated confidence, and prior art from 150+ community projects.

A curated collection of datasets for Large Language Models (LLMs), covering medical AI, NLP, multimodal learning, instruction tuning, reasoning, code generation, and evaluation benchmarks.

Benchmarking models as runtime Controllers for Loop Engineering

HiveMind Protocol - A Local-First, Privacy-Preserving Architecture for Agentic RAG

Research and writing can be automated. Authorship cannot. Evidence list + agent skills for spotting AI-produced papers no human stood behind.

211 free visual cards about large language models, from tokenization to deployment. Full collection: 376 cards.

The Best TypeScript framework for fine-tuning

Live truecolor terminal dashboard for a locally running LLM: tok/s, TTFT, GPUs, context, MTP acceptance, layer and real expert routing (llama.cpp)

MoE expert offload for low-VRAM GPUs — run 100B+ MoE models (DeepSeek, Qwen, Mixtral) on 8 GB cards. Expert cache with LRU hot experts, OpenAI-compatible proxy, GGUF multi-shard.

An experimental local AI runtime. It relays one request through five small models with one job each, on a single 8 GB consumer GPU.

Legend of Elya — N64 game with a real 6.36M-parameter ternary transformer on the VR4300 MIPS III CPU. Zelda-style dungeon, AI NPCs, byte-level inference at 1.23 tok/s scalar / 2.19 tok/s on the RSP overlay (measured under ares, never on silicon). Built with libdragon.

Ultimate Claude Fable 5 Guide 2026: Use Cases, Integrations & Benchmarks

[CVPR 2026] Official release of "Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning"