https://www.darius.wiki/en/topics/open-models/Open weights, licenses and small models you can run locallyenFri, 09 Oct 2026 00:00:00 GMThttps://www.darius.wiki/en/daily/2026-10-09/lightonocr-3/https://www.darius.wiki/en/daily/2026-10-09/lightonocr-3/Fri, 09 Oct 2026 00:00:00 GMTOpen sourceSmall document-parsing models with a new grounding mode that returns coordinates, suited to turning PDFs and scans into structured text. Take: The 0.8B is less than a point behind the 4B, which makes it attractive for document parsing in AI Cloud Drive: I'll run a comparison on the scans and table-heavy PDFs users upload most, to see whether some parsing can move from a cloud API onto our own machines.https://www.darius.wiki/en/daily/2026-10-09/liquid-ai-d1/https://www.darius.wiki/en/daily/2026-10-09/liquid-ai-d1/Fri, 09 Oct 2026 00:00:00 GMTOpen sourceSmall models that answer structured questions instead of generating text, aimed at devices and the edge. Take: Yes-or-no calls like moderation, routing or whether to hand off to a human don't need a generative model every time. I'll try d1-3B on PandaClaws' pre-publish compliance checks, but only as an evaluation until the license is clear.https://www.darius.wiki/en/daily/2026-10-09/nvidia-nemotron-ioi-imo/https://www.darius.wiki/en/daily/2026-10-09/nvidia-nemotron-ioi-imo/Fri, 09 Oct 2026 00:00:00 GMTResearchOne open model family clears the gold bar in both programming and math olympiads, with the data and evaluation pipelines released too. Take: Most teams won't run a 550B model; the part worth taking is the method: on IOI 2025 the Nano model (30B total, 3B active) went from 280 after SFT to 468 with test-time strategies like GenCorrect. If you build coding products, try a generate-then-self-correct layer on your own eval set first.https://www.darius.wiki/en/daily/2026-10-07/mistral-large-4-preview/https://www.darius.wiki/en/daily/2026-10-07/mistral-large-4-preview/Wed, 07 Oct 2026 00:00:00 GMTModel releaseMistral's largest model to date, natively multimodal; the preview API is live on Mistral Studio and the weights are due by the end of the month. Take: At $1.36 in and $4.18 out, it's worth running PandaClaws' long-form generation and multilingual rewriting against it. But this is a preview with no weights or license yet, so I'd benchmark it on my own eval set through the API now and only weigh self-hosting once the weights and license terms land at month end.https://www.darius.wiki/en/daily/2026-10-07/embeddinggemma-2/https://www.darius.wiki/en/daily/2026-10-07/embeddinggemma-2/Wed, 07 Oct 2026 00:00:00 GMTOpen sourceA small embedding model built on Gemma 4 that maps text, code, images, video and audio into one 768-dimensional space, bringing multimodal semantic search on device. Take: This one maps straight onto my AI Cloud Drive: with images, video, recordings and documents in one vector space, a single query can search across file types without a separate model per format. I'd swap the 270M text version in against my current embedder first, then load the vision module as needed; truncating to 128 dimensions cuts index storage, but measure the recall loss on your own retrieval set first.https://www.darius.wiki/en/daily/2026-10-06/aleph-alpha-kolibri-1/https://www.darius.wiki/en/daily/2026-10-06/aleph-alpha-kolibri-1/Tue, 06 Oct 2026 00:00:00 GMTOpen sourceAn open German-and-English model with a reasoning mode and tool calling; its FP8 weights take about 78 GB and serve on a single H200 or B200. Take: About 3B active per token, one GPU, Apache 2.0: this is the shape of model that suits privately deployed long-document Q&A like my AI Cloud Drive. But it only covers German and English, so don't drop it into Chinese workloads; what I want from it is a real throughput number for sliding-window attention plus long context on my own hardware before choosing the next self-hosted model.https://www.darius.wiki/en/daily/2026-10-05/allenai-astabrief-8b/https://www.darius.wiki/en/daily/2026-10-05/allenai-astabrief-8b/Mon, 05 Oct 2026 00:00:00 GMTOpen sourceAn 8B open-weights model turns retrieved excerpts into a cited long-form report, with weights and training data released together. Take: Retrieve-then-write-a-cited-answer is the most common request in my AI Cloud Drive's document Q&A. An 8B model that does it in one call means private deployment and a much lower bill. I'll compare citation accuracy on my own documents before replacing the large-model call.