https://oraculum.sh/news/Public model releases, updates, and open-source work from Oraculum, LLC.enSat, 26 Sep 2026 12:00:00 GMThttps://oraculum.sh/models/manchego/#servingoraculum-news-manchego-serve-manchego-offlineWed, 23 Sep 2026 12:00:00 GMTA small HTTP server that serves exactly the policy behind the model card’s numbers, with pinned, hash-verified weights and no outbound connections. On an NVIDIA A10: 81 ms median per question. Source: https://github.com/nschlaepfer/manchego-serveOpen sourcehttps://oraculum.sh/models/manchego/#evaluationoraculum-news-manchego-against-the-whole-fieldWed, 23 Sep 2026 12:00:00 GMTThe Manchego cards now compare v2.1 with every trained open ~4B decision model that could be run as shipped (28 in all, counting v2.1), and disclose how its training data was designed, audited and selected. Source: https://huggingface.co/oraculumai/ManchegoModel updatehttps://oraculum.sh/models/manchego-mlx-8bit/oraculum-news-manchego-on-apple-siliconWed, 23 Sep 2026 12:00:00 GMTThe MLX 8-bit and 4-bit builds of v2.1, with format results on their cards. Eight bits stays within one decision of bf16; four bits costs six of 80 held-out rows, and says so. Source: https://huggingface.co/oraculumai/Manchego-MLX-8bitModel updatehttps://oraculum.sh/models/manchego/oraculum-news-meet-manchego-v2-1Mon, 21 Sep 2026 12:00:00 GMTA second training stage on human-written task definitions repairs v2’s overconfidence on unfamiliar tasks. Calibration error on a sealed set of 25 unseen tasks: 0.153 to 0.040. Source: https://huggingface.co/oraculumai/ManchegoReleasehttps://oraculum.sh/models/manchego/#versionsoraculum-news-manchego-v2Sun, 20 Sep 2026 12:00:00 GMTTrained from the base on 491,520 rows with computed labels and original human annotations. It remains available at the tag v2. Source: https://huggingface.co/oraculumai/ManchegoReleasehttps://oraculum.sh/models/manchego/#versionsoraculum-news-manchego-v1-the-first-releaseSat, 19 Sep 2026 12:00:00 GMTThe first release of the 4B decision model. It remains available at the tag v1. Source: https://huggingface.co/oraculumai/ManchegoReleasehttps://oraculum.sh/models/cbramod-mi/oraculum-news-motor-imagery-with-a-clearThu, 20 Aug 2026 12:00:00 GMTA cue-paced, 14-channel Core ML classifier: 0.784 on unseen users under subject-grouped cross-validation, 210 of 210 decisions preserved in the port. Source: https://huggingface.co/oraculumai/CBraMod-MI-CoreMLModel updatehttps://oraculum.sh/models/cbramod-embeddings/oraculum-news-eeg-representations-for-apple-platformsWed, 19 Aug 2026 12:00:00 GMTThe pretrained CBraMod backbone as a Core ML embedder, matching PyTorch to a worst relative L2 of 1.3 × 10⁻⁵ on real EEG. Source: https://huggingface.co/oraculumai/CBraMod-CoreML-AppleModel updatehttps://oraculum.sh/models/zuna-1-1/#upstream-fixoraculum-news-a-numerical-fix-for-coreTue, 18 Aug 2026 12:00:00 GMTPorting ZUNA1.1 surfaced an epsilon-inflation bug in coremltools’ rms_norm conversion, large enough to change outputs by whole percents at fp32. Reported upstream with a proposed fix and regression tests. Source: https://github.com/apple/coremltools/pull/2822Open sourcehttps://oraculum.sh/models/zuna-1-1/oraculum-news-zuna1-1-on-apple-platformsTue, 18 Aug 2026 12:00:00 GMTZyphra’s newer EEG reconstruction model as fp32 Core ML profiles, with a 14-channel default for EPOC X-style headsets. Source: https://huggingface.co/oraculumai/ZUNA1.1-CoreML-AppleModel update