DIY Generative AI

1.1 Image generation

When models like DALL-E first appeared, they lived in a corporate walled garden. The 2022 release of Stable Diffusion opened the field — it was a potent, open-weights model that let anyone with a decent gaming PC generate high-quality images. Compute came from Stability AI; the architecture and training methodology were published as papers; the dataset came from LAION, a non-profit that had done the hard work of scraping and aligning image-text pairs at scale; inference clients sprang up from the community in short order. All four ingredients in public (although I suspect the training story and the data story might have been selectively curated).

And then we got the other thing that open-source communities excel in: Drama!

In 2024 a chunk of the original Stable Diffusion researchers walked out of Stability AI over disagreements about open-source commitments and commercialization strategy, regrouped as Black Forest Labs, and shipped the Flux family — a transformer-based model released as a “dev” version (open-weights, non-commercial) and a “pro” version (paid API). Their editing variant, FLUX.1 Kontext, was AFAICT the first open-weights edit model of professional capability.

The open/commercial friction is ongoing: corporate models (Flux pro, DALL-E 3, Midjourney) offer plug-and-play polish but lock advanced features behind APIs and are not very customizable; community-tuned ones (SDXL, the CivitAI menagerie) offer customization, autonomy, and the ability to coax out things the mainstream models’ guardrails won’t — idiosyncratic styles, violence, sexual content, nazis.

1.2 Language model pirates

When GPT-3 was kept under lock and key, a scrappy crew of volunteer researchers on a Discord server formed EleutherAI in 2020. Their explicit goal was to replicate the model and give it away. They bootstrapped their own massive dataset, The Pile, released GPT-Neo, GPT-J, and GPT-NeoX-20B in turn, and demonstrated that a distributed collective could take on the giants (at least, if a giant game them some free compute).

Meta later started open-weights releases of more powerful models, beginning with LLaMA (initially leaked rather than released; Meta has since seemingly embraced the leak as strategy), shifting the open-weights frontier into the multi-billion-parameter range. AFAICT EleutherAI got the open-LLM ball rolling, Meta made it big, Mistral added a European actor — and then…

1.3 Chinese open-weights

In 2024, Chinese labs decided there was no particular reason to follow the script. Alibaba’s Qwen series, DeepSeek-V2/V3 and the R1 reasoning model, Moonshot’s Kimi, Zhipu’s GLM, 01.AI’s Yi — and a steady drip of smaller releases from Tencent (Hunyuan), ByteDance, MiniMax, StepFun, and others — arrived with permissive licences, fluent in English and Chinese, at scales and quality matching or beating closed Western models on many benchmarks, maybe.

Given US export controls on high-end GPUs, aggressive open-sourcing has been widely read as a baller geopolitical play as much as a research strategy — recruit downstream users, distribute the costs of fine-tuning and tooling, build legitimacy outside the export-control-and-API regime, commoditize the layer above whoever is selling the chips.

DeepSeek’s R1 in January 2025 was the first to upset the apple cart — an open-weights reasoning model competitive with OpenAI’s o1, released with a paper explaining the RL training recipe in unusual detail. It cratered Nvidia’s market cap for a week. Whether the architectural and training-cost claims fully replicate or not¹, the availability changed expectations. “Open” no longer means “a year behind”.

These labs differ in how much they reveal upstream of the weights. DeepSeek seems to publish more methodology and partial training detail than most; the rest release weights, a model card, and not much more. Training data: essentially never. Same conflation as everywhere else — but the closed-model players seemingly cannot use “we’re so far ahead it doesn’t matter” as a rebuttal any more.

There are roadblocks to western use. Chinese regulation requires these models to deflect on politically sensitive topics, so the weights arrive pre-aligned to Beijing. That deflection is, at least for now, vulnerable to abliteration, so it is relatively easy for the community to remove it — uncensored DeepSeek and Qwen variants are a minor cottage industry.

1.4 Sovereign and public-good initiatives

Buzzword: “Middle powers”.

A small-but-interesting cluster of state-funded open-weights efforts exists. Distinct from the US-commercial and Chinese-commercial models, these are non-superpower governments that decided AI sovereign capacity was strategically worth doing publicly. Switzerland’s Apertus (ETH, EPFL, and the Swiss National Supercomputing Centre) publishes their whole training recipe — weights, training data, methods — built for the public good with serious effort on Swiss German, Romansh, and other long-underserved languages. Singapore’s SEA-LION does something similar for 11 Southeast Asian languages — Thai, Burmese, Lao, Khmer, Tamil, Tagalog, and others the global frontier models handle poorly. The UAE’s Falcon family from the UAE fits the same pattern, and France’s Jean Zay supercomputer was behind BigScience’s BLOOM. These projects usually trail the commercial frontier but encode different priorities — multilingualism, regional language coverage, public-interest licensing.

1.5 Hacking on a shoestring

Buzzword: Jugaad, a kind of frugal, resourceful, get-it-done innovation that thrives in the absence of capital-rich environments by making do with what’s at hand, and being unfraid to make things that will fall apart relatively quickly. Make do, hack together what’s at hand.

Especially hot with respect to globalizing AI, breaking it free from geographically concentrated power. One poster child for Jugaad is Masakhane (“we build together”), a grassroots, pan-African collective building NLP models for African languages insufficiently easy/profitable to be served by the commercial labs.

1.6 Community projects

The Chinese open-weights wave buys us the ability to host some inference in out own geographic region. I’ve been costing out a 50-person Australian compute club that would run such models on our own hardware. Follow along at danmackinlay/SOV.

1.7 Fandoms

A reliable driver for weird obsession-heavy project is… nerds who like weird stuff. Fandoms are tried-and-tested distributed innovation communities, and even for AI this is a thing.

First data point: the My Little Pony fandom. When Friendship is Magic ended in 2019, a segment of the fanbase decided they weren’t done with it yet. This led to the Pony Preservation Project, which originated on 4chan’s /mlp/ board with a mission to meticulously collect, clean, and align voice data from every episode, building a high-quality dataset specifically for AI training. That dataset became a key component for 15.ai, a landmark text-to-speech platform that launched in 2020 and let anyone speak pony voices. That led to much fun and equally much drama. The result was an explosion of fan-made content, including fully AI-voiced animations, effectively continuing the show through sheer force of will and technical ingenuity.

Second data point: AI HUB. AI HUB became the go-to Discord community for the ethically murky world of cloning the voices of actual musicians and other public figures. The community was a hotbed for sharing and developing voice-cloning techniques themselves — pre-trained models of specific singers, techniques for separating vocals from tracks to make clean training data, training pipelines, sometimes even algorithms. It was a volunteer-run decentralised R&D and distribution network that lowers the barrier to entry for a controversial and technically demanding application of AI.

I use the past tense: it it’s not quite as active as it was in 2024 after dispersing a little under he bullwhip of the copyright lawyers. Fan projects frequently test the edges of copyright law, and fans often have smaller budgets than copyright holders or large tech companies, so these projects either get mired in drama or keep a low profile.

Porn is probably in the same category, but I bet they have more money behind them.

1.8 Mandatory AI safety qualm

As with all DIY, this stuff is dangerous. Insofar as empowering people with more tools makes people more powerful, democratizing it puts it in the hands both of nice, disempowered people and of terrorists. Except these machines might themselves eventually autonomously behave nicely, or terroristically.

OTOH, there is also a danger in letting powerful models be in the hands of only a small number of geopolitical actors. YMMV

2.1 Hugging Face

A central repository for much of open-source AI — models, datasets, and the transformers library that lets us actually use the stuff without pulling our hair out. AFAICT every other organization on this list, from EleutherAI to DeepSeek, hosts at least some models there.

2.2 CivitAI

CivitAI is a specialised marketplace for the diffusion-model fine-tuning community. It hosts the LoRAs, textual inversions, full fine-tunes, specialize image generation on niche styles — anime, photorealism, particular painters, particular fictional characters, particular kinks. Visual browsing with prompt-and-output previews; simple download; an API that the local diffusion clients (DiffusionBee, Draw Things, ComfyUI, …) plug into. More permissive content policy than Hugging Face. Pony Diffusion V6 XL — an SDXL fine-tune trained on character art from booru-style databases, with no RLHF passes applied — is a concrete example of the community making explicit choices against the alignment priorities of production models. The downstream ecosystem of hundreds of derivative fine-tunes if fertile with dark, surreal and thirsty image generation. See the project about page.

2.3 Allen Institute for AI (AI2)

Bankrolled by the late Microsoft co-founder Paul Allen. AI2 is a non-profit that tackles long-term research that corporations apparently won’t and universities seemingly can’t. They built foundational tools like AllenNLP for language research and the Semantic Scholar search engine. With their OLMo model they released not just the weights but the training data and code as well.

2.4 EleutherAI

A grassroots collective born on Discord around 2020 that replicated GPT-3-class models and gave them away. They scraped the web to assemble The Pile dataset when, AFAICT, nobody else was doing it openly at that scale. They continue to produce interesting research infrastructure and weird models, and publish interesting results.

2.5 LAION

A German non-profit of data-hoarders-for-the-people, with a mission to create and release the massive datasets that AI models need to learn from. LAION-5B (5.8 billion image-text pairs) is their best-known release. LAION has leveled up the image gen data sets. Without their work, there would be no Stable Diffusion etc. They are upstream of much of the open image-generation pipeline.

2.6 BigScience

Coordinated by Hugging Face, it brought over 1,000 researchers together in a global jam session to build BLOOM, a massive multilingual language model. It aimed to be a proof of concept for a different way of doing research — open, global, collaborative — and a counter-narrative to the idea that only secretive corporate labs could build state-of-the-art models. AFAICT nobody has repeated it at the same scale since.

2.7 Chinese open-weights labs

As of mid-2026 the roster included DeepSeek on coding, math, and reasoning. Qwen (Alibaba) does broad-purpose multilingual and is apparently the most-downloaded open-weights family on HuggingFace by some margin. Kimi (Moonshot) covers long-context and agentic tasks. GLM (Zhipu, out of Tsinghua’s KEG lab) covers bilingual chat. Yi (01.AI, Kai-Fu Lee’s outfit) does efficient general-purpose models. Thereafter can be found a long tail of smaller releases from Tencent, ByteDance, MiniMax, StepFun, and others. The pecking order shifts every quarter.

These various labs have made themselves the protagonists of open-source AI drama. In 2026, plot beats were played by Chinese commercial labs with strategic reasons to give weights away. Intersting incentive/market/geopolitcal structures in play. — export-control geopolitics, market structure (no domestic OpenAI-equivalent to undercut). Releases are mostly weights-plus-model-card; DeepSeek seems to be the outlier in publishing methodology; training data is essentially never released by anyone.

2.8 DAIR Institute

Founded by Dr. Timnit Gebru after her high-profile exit from Google. An independent research institute. “We publish interdisciplinary work uncovering and mitigating the harms of current AI systems, and research, tools and frameworks for the technological future we should build instead.” It aspires to be a ‘conscience’ of the scene, in the sense of speaking up about present harms.

2.9 Agora

A hyper-grassroots collective of engineers and creators organised on Discord Focuses on bleeding edge hip thing — multi-agent systems, real-time learning, augmenting AI reasoning — and radically open in the share-before-it-works mold. There are surely are other projects like Agora; it has the distinction of being the one I encountered.

2.10 Oumi

Oumi is Python library/platform convetion for the pipelines of data prep and training to DIY models from scratch. ## Democratizing training on consumer hardware

See distributed NN training.