Most of the interesting open-source LLM work right now isn't a new model. It's the plumbing around models: the pipelines that fine-tune them, the bots that put them in a group chat, the tools that turn a video into subtitles in another language. I went back through my bookmarked notes on that layer and pulled out eight projects that, between them, cover most of the path from training a model to putting it in front of people. My view after rereading them: the hard part has moved from "can a model do this" to "who owns the glue," and the projects below are where that glue is being written in the open.
Training and evaluating
Oumi. If you want one toolchain for data prep, fine-tuning, evaluation and serving, Oumi is the most complete open option I've saved. It handles SFT, LoRA/QLoRA, DPO and GRPO across text and vision models (Qwen, Gemma, Llama, DeepSeek, GLM and more), runs from a laptop up to cloud clusters via oumi launch, and hands inference to vLLM or SGLang. The recent releases lean hard into agents, with tool-calling SFT/DPO, executable tool environments and rubric-based judges. My notes cover Oumi's end-to-end training and deployment workflow in more detail.
How small can an agent framework be? A tutorial built on PocketFlow argues that every LLM agent reduces to a graph: a decision node, some action nodes, an end node, and a loop back to the decision. PocketFlow does it in roughly 100 lines, with each node split into prep, exec and post steps over a shared store. The worked example is a research assistant that decides whether to search or answer, and the same shape turns out to be hiding inside LangGraph, the OpenAI Agents SDK and Pydantic's agents. It's a useful antidote to framework fatigue: why agents are just graphs with loops.
Voice and video
F5-TTS. Flow matching has become a serious alternative to autoregressive speech synthesis, and F5-TTS is the reference implementation people keep pointing to. It pairs a Diffusion Transformer with ConvNeXt V2, adds a "Sway Sampling" trick at inference time, installs with a single pip install f5-tts, and ships a Gradio app for multi-speaker generation and voice chat. One thing to check before shipping anything: the code is MIT, but the pretrained weights are CC-BY-NC because of their training data. Details are in my write-up on F5-TTS and its licensing split.
VideoLingo. This is the project that made me realise how much of "AI dubbing" is pipeline engineering. VideoLingo chains download, word-level ASR and alignment, subtitle segmentation, terminology-aware translation with an optional reflection pass, and dubbing through a choice of TTS backends (F5-TTS among them), all behind a Streamlit UI or a local HTTP API. The note is candid about where it breaks: background noise, mixed-language speech and strict JSON output from the LLM. See how VideoLingo turns a video into translated, dubbed subtitles.
Putting models where people already talk
LangBot. Instead of building a chat UI, LangBot connects models to the chat apps people already use: Discord, Telegram, Slack, LINE, QQ, WeChat, Feishu, DingTalk and more, plus a Matrix bridge for the long tail. It brings the boring-but-necessary production pieces (access control, rate limits, content filtering, monitoring), a plugin system with MCP support, and hooks into Dify, Coze and n8n. You can try it with uvx langbot and a browser. I summarised LangBot's platform coverage and plugin model.
HuixiangDou. Group chats are a hostile place for a bot: most messages aren't questions, and answering everything is spam. HuixiangDou handles this with a three-stage pipeline (preprocess, reject, respond) so it stays quiet unless it has something grounded to say. It needs no training, runs in CPU-only, 2 GB and 10 GB GPU configurations, and tunes its scope through lists of good and bad example questions. The successor, HuixiangDou2, moves to GraphRAG. Here's what HuixiangDou does differently in group chats.
Chatwoot. When the conversation is with customers rather than colleagues, you need an inbox, not a bot. Chatwoot is an MIT-licensed, self-hostable alternative to Intercom and Zendesk that pulls live chat, email, WhatsApp, Telegram, Instagram and SMS into one place, and adds an AI agent called Captain to deflect routine questions. It also covers help-center articles, CSAT reports and Shopify order lookup. My notes on running Chatwoot as an open-source support desk list the deployment options.
A small personal use case
LLM-summarised bookmarks. Not every LLM application needs a platform. One blogger wired a bookmark extension, GitHub Actions, the Wayback Machine, Jina Reader and a cheap DeepSeek model into a pipeline that saves each bookmarked page's full text and writes a list summary plus a one-line gist, for under one yuan a month. It fixes two real problems at once: dead links and forgetting what a long article said. Read the bookmark summary and full-text index setup.
Taken together, these eight sketch a stack you can assemble without a single proprietary service: train or adapt a model with Oumi, keep the agent logic as a plain graph, give it a voice with F5-TTS, localise video with VideoLingo, and meet users in chat through LangBot, HuixiangDou or Chatwoot. Licences are the thing to watch, more than capability. If you've put one of these into production, I'd like to hear which piece needed the most glue.
Top comments (0)