Release
Manchego V3: a fresh decision model.
A new run from Qwen3.5-4B under the SemIf prompt, with public benchmark gains, a documented out-of-distribution regression, and full calibration limits.
News & releases
The latest signals.
Releases, model-card updates, and open-source work, each linked to its public source.
All entries
Model update
V3, in two MLX formats.
The 4.5 GB 8-bit and 2.4 GB 4-bit builds bring the V3 SemIf contract to Apple silicon. Each format publishes its own conversion checks.
Open source
manchego-serve 0.2: versioned contracts.
V3 and V2.1 keep their own prompt and serving policy. Per-question contracts, temperatures and weight verification are reported with the response.
Open source
manchego-serve: Manchego, offline.
The V2.1 server release used pinned, hash-verified weights and no outbound requests. Its published NVIDIA A10 check measured 81 ms median per question; that result is historical, not V3 latency.
Model update
Manchego, against the whole field.
The V2.1 cards compared v2.1 with every trained open ~4B decision model that could be run as shipped (28 in all, counting v2.1), and disclosed how its training data was designed, audited and selected.
Model update
Manchego, on Apple silicon.
The MLX 8-bit and 4-bit builds of v2.1, with format results on their cards. The historical format check put eight bits within one decision of bf16 and four bits six decisions lower on 80 rows.
Release
Meet Manchego v2.1.
A 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.
Release
Manchego v2.
Trained from the base on 491,520 rows with computed labels and original human annotations. A historical research iteration, superseded by later releases.
Release
Manchego v1, the first release.
The first release of the 4B decision model. A historical research iteration, superseded by later releases.
Model update
Motor imagery, with a clear contract.
A 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.
Model update
EEG representations for Apple platforms.
The pretrained CBraMod backbone as a Core ML embedder, matching PyTorch to a worst relative L2 of 1.3 × 10⁻⁵ on real EEG.
Open source
A numerical fix for Core ML.
Porting 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.
Model update
ZUNA1.1, on Apple platforms.
Zyphra’s newer EEG reconstruction model as fp32 Core ML profiles, with a 14-channel default for EPOC X-style headsets.
Dates are repository updates unless a tagged release date is stated. Each entry links to its public source.
Follow the work
Good questions. Measured answers.
New models, their evidence, and the occasional correction. Model cards come first; this site follows.