Data stories from the AI hub

Of the 10,491 open-source AI projects in the olud.ai catalogue, 4,436 — 42.3% — have had no push in six months. Even among the 1,150 repos with 10,000+ stars, 17.1% have gone quiet.

Across 270 AI models with both a benchmark score and a published price, the best open model scores 59.7 against 63.1 for the best closed one. Per point of intelligence, open models cost 3.1× less.

A 24 GB card runs a 35B model, and an 8 GB card runs Qwen3 8B in an estimated 6.4 GB. The honest VRAM table, built on the published rule of 0.58 GB per billion parameters — margin ±40%.

We mapped the 700 MCP servers in the olud.ai hub: 36.4% fit no established category, 22.4% are aggregators, and the largest single-purpose category is Memory & Knowledge — agents are getting memories before hands.

LanguageTool, Harper and Vale are three open-source alternatives to Grammarly, all of which shipped code this week. Harper is maintained by Automattic and runs entirely offline — nothing you type leaves your machine.

50 of the 76 closed models we price and score are outperformed by another model from the same company that also costs less — on input and on output. The worst gap is 20×.

1,641 of the 8,246 AI software repositories we track carry no identifiable licence, and together they hold 8.3 million stars. The rate barely moves between a 200-star experiment and a 50,000-star tool.

41 of the 100 most-starred AI projects were created in 2025 or 2026, and eighteen of them did not exist in January. Yet not one of the top 50 has stopped receiving commits — the ranking turns over through replacement, not decay.

97 of the 190 commercially priced models we track are Chinese, and Alibaba alone ships 49 of them. They are not, however, the cheap ones — the median Chinese model costs nearly four times the median American one.

Daily price tracking of 63 models: 34 open-weight models got more expensive, 13 got cheaper. GLM 5.2 changed price 22 times in 25 days — and nobody at Z.ai decided any of it.

Across 10,216 repositories grouped by the year they were created, Python held between 53 % and 57 % of every cohort from 2017 to 2024 — then lost twenty-one points. TypeScript is now at 27 %.

223 of 848 AI code repositories above 10,000 stars have gone a month without a commit, and 63 have gone a full year. Star counts do not decay, so nothing on the page tells you.

Claude Opus 5, Claude Fable 5, GPT-5.6 and Kimi K3 land within 3.6 points of each other on intelligence while their prices vary by 3.3×. Price no longer tracks capability at the top.

Claude Opus 5 shipped forty-five days after Claude Fable 5. It scores higher on both composite indices and costs exactly half as much per token. That should not happen, and the reason it does is more interesting than the ranking.

It scores 57.1 on intelligence at $15 per million output tokens, ahead of GPT-5.6 Terra at the same price and 1.8 points behind GPT-5.6 Sol, which costs twice as much. Then you read the parameter count: 2.8 trillion.

Sol, Terra and Luna shipped on the same day at $30, $15 and $6 per million output tokens. At maximum effort they sit 7.7 intelligence points apart. The six effort levels inside Luna alone sit 24.6 points apart.

We tested ChatGPT (GPT-4o) against Llama 3.3, DeepSeek R1, and Mistral on 6 real-world tasks. Honest benchmarks, clear verdict. Are free open-source models finally good enough?

8 open-source AI coding assistants tested in 2026 — OpenCode, Cline, Aider, Continue, Tabby, Void and more. Free, self-hostable GitHub Copilot alternatives, plus the best local coding…

6 free open-source AI image generators tested in 2026 - ComfyUI, AUTOMATIC1111, Fooocus, InvokeAI and more - plus the best models (FLUX, Stable Diffusion, Sana). Self-hostable Midjourney…

A simple 2026 guide to which local AI model fits your GPU - 8GB, 12GB, 16GB, 24GB and beyond. The VRAM rule of thumb, model picks per tier, plus Apple Silicon and CPU.

An honest 2026 comparison of the four leading open-weight AI families - DeepSeek, Llama, Qwen and Mistral - by license, hardware needs, coding, reasoning and which to choose.

Kimi K2.6, DeepSeek V4, GLM-5.1, Llama 4, Qwen, Gemma 4 and Mistral — ranked on reasoning, coding, context and price. Which one to pick, and why.

No subscription, no cloud, fully private. Hardware by budget, the best free tools (Ollama, LM Studio, Jan), and exactly which model fits your machine.

Real benchmarks, full feature comparison table, and a clear verdict. Ollama for developers, LM Studio for everyone else.

Ollama, ComfyUI, Whisper, LangChain — the complete guide to the best open-source AI tools, tested and ranked.

Writers, podcasters, designers, video creators — everything you need to create with AI for free.

DeepSeek, Llama, Qwen, Mistral and more — private, self-hostable models to use instead of ChatGPT, with live pricing and a local setup guide.

Open-weight models with strong reasoning and long context — DeepSeek, Qwen, GLM and more you can self-host instead of Claude.

Multimodal, long-context open models like Gemma, Qwen and Kimi — self-hostable replacements for Google Gemini.

Flux, Stable Diffusion, ComfyUI and Fooocus — generate AI images on your own machine, for free, with full control.

Continue, Aider, Tabby and open coding models — AI completions in your editor, optionally fully offline and private.

Perplexica (Vane), Morphic and Khoj — self-host private AI search with web citations, using local or hosted models.

Our full directory of open-source, self-hostable replacements for the most popular closed AI products — chat, images, coding and search.