# 天工 / Tiangong

> 本地大模型机器人玩法汇总站
> The hub for robot and AI plays powered by local LLMs

Bilingual site (Chinese at /, English at /en/). Every article has a Markdown version (URL ending in .md).
Full text of all articles: https://spark.tlec.top/llms-full.txt

## 玩法收藏 (zh)

- [让你的 Agent 把收藏整理成天工玩法](https://spark.tlec.top/plays/agent-content-publishing.md): 从链接和部署笔记开始，通过内容接口保存草稿、审核、发布，并自动更新网页和搜索。
- [用 Ollama + Open WebUI 搭一个家用 ChatGPT](https://spark.tlec.top/plays/ollama-open-webui.md): 十分钟在自己电脑上跑起带网页界面、多用户、可切换模型的聊天助手，局域网里的手机也能用。
- [在 VS Code 里接一个完全离线的编程助手](https://spark.tlec.top/plays/local-coding-assistant.md): 用 Continue 插件连接本地 Ollama 模型，补全和对话都不出本机，适合公司代码不能外传的场景。
- [用 AnythingLLM 把自己的文档变成可问答的知识库](https://spark.tlec.top/plays/anythingllm-knowledge-base.md): 桌面版一键安装，接上本地模型和本地向量库，拖进 PDF、网页、笔记就能问。

## 博客 (zh)

- [为什么要做这个站](https://spark.tlec.top/blog/hello.md): 本地模型越来越好用，但玩法散落在各个论坛和群里，想找个地方把它们攒起来。

## 资讯 (zh)

- [资讯栏目说明](https://spark.tlec.top/news/news-section-launch.md): 本栏目由 AI 每天从 Reddit、Hacker News、GitHub Releases 等来源收集本地模型部署相关资讯，筛选、摘要、翻译后经人工审核发布。
- [Gemma-4-31B-Humanlike-Chat 发布：优化人类语音与对话风格的微调模型](https://spark.tlec.top/news/gemma-4-31b-humanlike-chat-released-63cb1ecb.md): 基于 Google Gemma 4 31B 官方模型微调的 Humanlike-Chat 版本发布，改进了人类语音与聊天风格，支持 GGUF、Safetensors、FP8 及 GPTQ-Int4 本地部署权重。
- [Unsloth 发布 DeepSeek-V4.1-Flash 动态 v3.1 GGUF 量化模型](https://spark.tlec.top/news/introducing-deepseek-v41-flash-1017-dynamic-v31-unsloth-gguf-6b1222de.md): Unsloth 推出 DeepSeek-V4.1-1017 动态 v3.1 预览版 GGUF，在多项基准测试中展现出高精度保留能力，并更新了 Unsloth Desktop 功能。
- [ISTA 发布 DeepSeek-V4.1-Flash 量化版本基准测试说明](https://spark.tlec.top/news/ista-deepseek-v41-flash-1017-gsqrco-quants-comparison-bcca3978.md): ISTA 针对与 Unsloth 测试结果的差异做出说明，指出推理预算限制（32K vs 128K）是导致跑分差异的主因。
- [Google Project Suncatcher 卫星成功在太空本地运行 Gemma 模型](https://spark.tlec.top/news/local-ai-made-it-to-space-gemma-tpu-66b99d03.md): Google 确认其 Project Suncatcher 卫星成功在板载 TPU 上运行 Gemma 模型，并向地球传回了生成的首个响应（一首诗）。
- [MOSS 团队 Reddit AMA：开源模型家族全貌与本地运行资源汇总](https://spark.tlec.top/news/moss-team-ama-local-llama-2b1877ce.md): 复旦大学及上海创新研究院背景的 MOSS 团队在 Reddit 进行 AMA，集中介绍了涵盖图像生成、语音合成、音频与视觉理解的开源模型家族（如 MOSS-Image、MOSS-TTS 系列等）。
- [商汤 SenseNova-Loop-9B-Preview 多模态开源模型发布](https://spark.tlec.top/news/sensenova-loop-9b-preview-40479274.md): SenseNova-Loop-9B-Preview 是一款面向长文本和 Agent 工作流的多模态语言模型，采用共享循环解码器和双流流形约束超连接（mHC）架构，支持 256K 超长上下文。
- [Sketch Stadium：基于 MCP 与推送通知让 AI Agent 玩“你画我猜”](https://spark.tlec.top/news/sketch-stadium-ai-agent-pictionary-ec56c328.md): 开发者推出 Sketch Stadium 演示项目，让 AI Agent 通过 SVG 绘制和猜测来玩 Pictionary（你画我猜），同时展示了 Wakeshift 推送通知 SDK 与 MCP 的结合应用。
- [llama.cpp 升级 DFlash2：Qwen3.8-Flash-Next 解码速度平均提升 17%](https://spark.tlec.top/news/upgrade-flash-next-to-dflash2-6ac0e32f.md): 随着 llama.cpp 对 DFlash2 的支持，Qwen3.8-Flash-Next 模型在本地运行的解码速度平均提升 17%，草稿头体积大幅缩减至约 350MB。

## Plays (en)

- [Let your Agent publish local-model plays to Tiangong](https://spark.tlec.top/en/plays/agent-content-publishing.md): Turn links and deployment notes into reviewed articles through the Tiangong content API.
- [A home ChatGPT with Ollama + Open WebUI](https://spark.tlec.top/en/plays/ollama-open-webui.md): Get a multi-user, model-switching chat assistant with a web UI running on your own machine in ten minutes — usable from any phone on your LAN.
- [A fully offline coding assistant in VS Code](https://spark.tlec.top/en/plays/local-coding-assistant.md): Connect the Continue extension to a local Ollama model so completions and chat never leave your machine — great when company code can't go to the cloud.
- [Turn your documents into a Q&A knowledge base with AnythingLLM](https://spark.tlec.top/en/plays/anythingllm-knowledge-base.md): One-click desktop install, wired to a local model and a local vector store — drop in PDFs, web pages and notes and start asking.

## Blog (en)

- [Why this site exists](https://spark.tlec.top/en/blog/hello.md): Local models keep getting better, but the fun ways to use them are scattered across forums and group chats. I wanted one place to collect them.

## News (en)

- [About the news section](https://spark.tlec.top/en/news/news-section-launch.md): AI collects local-LLM deployment news daily from Reddit, Hacker News, GitHub Releases and more; items are filtered, summarized, translated and reviewed by a human before publishing.
- [Gemma-4-31B-Humanlike-Chat Released: Fine-Tuned for Natural Human-Like Voice](https://spark.tlec.top/en/news/gemma-4-31b-humanlike-chat-released-63cb1ecb.md): A fine-tuned version of Google's Gemma 4 31B optimized for natural human conversation style and voice, with GGUF, Safetensors, FP8, and GPTQ-Int4 weights available for local deployment.
- [Unsloth Releases DeepSeek-V4.1-Flash Dynamic v3.1 GGUFs](https://spark.tlec.top/en/news/introducing-deepseek-v41-flash-1017-dynamic-v31-unsloth-gguf-6b1222de.md): Unsloth introduces DeepSeek-V4.1-1017 Dynamic v3.1 preview GGUFs featuring high accuracy retention across benchmarks, alongside Unsloth Desktop updates.
- [ISTA DeepSeek-V4.1-Flash Quantization Benchmark Comparison Clarification](https://spark.tlec.top/en/news/ista-deepseek-v41-flash-1017-gsqrco-quants-comparison-bcca3978.md): ISTA addresses benchmark discrepancies with Unsloth regarding DeepSeek-V4.1-Flash-1017, highlighting the impact of reasoning budget limits.
- [Google's Project Suncatcher Satellite Successfully Runs Gemma Locally in Space](https://spark.tlec.top/en/news/local-ai-made-it-to-space-gemma-tpu-66b99d03.md): Google confirmed that its Project Suncatcher satellite successfully ran a Gemma model on an onboard TPU in space and transmitted its first generated response back to Earth.
- [MOSS Team Reddit AMA: Overview of Open Model Family and Local Resources](https://spark.tlec.top/en/news/moss-team-ama-local-llama-2b1877ce.md): The MOSS team from Fudan University and the Shanghai Innovation Institute hosted an AMA on Reddit, showcasing their open-source model family spanning image generation, speech synthesis, audio, and visual understanding.
- [SenseNova-Loop-9B-Preview Multimodal Model Released](https://spark.tlec.top/en/news/sensenova-loop-9b-preview-40479274.md): SenseNova-Loop-9B-Preview is a multimodal language model designed for long-context and agentic workflows, featuring a shared looped decoder and two-stream manifold-constrained hyper-connections (mHC).
- [Sketch Stadium: Let Your AI Agent Play Pictionary via MCP and Push Notifications](https://spark.tlec.top/en/news/sketch-stadium-ai-agent-pictionary-ec56c328.md): A new tech demo called Sketch Stadium allows AI agents to play Pictionary using a subset of SVG, showcasing the Wakeshift SDK for agent push notifications and MCP integration.
- [Upgrade Flash Next to DFlash2 in llama.cpp for ~17% Speed Boost](https://spark.tlec.top/en/news/upgrade-flash-next-to-dflash2-6ac0e32f.md): With llama.cpp now supporting DFlash2 for Qwen3.8-Flash-Next, local decode speeds improve by an average of 17% while draft head size shrinks to ~350MB.