oraculumHugging Face
Manchego v2.1 · calibrated on unseen tasks

An independent AI research lab

Intelligence, closer to the signal.

We build small decision models you can inspect, and bring EEG foundation models to Apple devices. Every release ships with its evidence and its limits.

Explore the researchMeet the models
Fig. 000Signal → Structure → Decision

A generative study. Not recorded EEG or model output.

8

releases in the catalog

Decision models, conversions, and EEG packages

0.805

JevBench accuracy

Manchego v2.1 on 231 public decisions: 3= of 28 open ~4B models

0.040

calibration error, unseen tasks

Down from 0.153 in v2, on a sealed set opened once

210/210

decisions preserved

Motor imagery, PyTorch to Core ML, on real EEG

01Selected work

Two lines of research. One standard of evidence.

How machines choose, and how neural signals become usable. Each release is small, specific, and measured.

Decision models · Manchego v2.1

A decision is
a distribution.

Manchego reads a state, a question and a closed set of options, and returns a probability for every option from one forward pass. It never writes a word, so it can tell you when it isn’t sure.

Read the model cardmanchego-serve

Fig. 01 · Recorded response

Question asked about the same message:

One message

Customerthe blender I bought last week smells of burning and stopped working. Order 5521.

Three separate questions about this message, each answered in one forward pass.

Question · Choice

Which team should handle this message?

3 options offered. It reads them all at once and writes nothing.

Probability of each optionchance, 1/3

  1. Areturns0.998
  2. Bshipping0.001
  3. Cbilling0.001

confidence0.998

Returns, almost certainly.The recorded answer, T = 1

The other 2 options share 0.002 between them.

Returned"route": {"choice": "returns", "confidence": 0.998}0 output tokens

A returns: 0.998; B shipping: 0.001; C billing: 0.001.

Question · Yes or no

Does the customer report a defective product?

2 options offered. It reads them all at once and writes nothing.

Probability of each optionchance, 1/2

  1. Ayes0.992
  2. Bno0.008

P(yes)0.992

Yes, at 0.992.The recorded answer, T = 1

A yes-or-no question is a two-option prompt; the answer is the probability of yes.

Returned"defective": {"noul": 0.992}0 output tokens

A yes: 0.992; B no: 0.008.

Question · Score

How urgent is this message?

3 options offered. It reads them all at once and writes nothing.

Probability of each optionchance, 1/3

  1. Aroutine level 00.104
  2. Bsoon level 10.454
  3. Cimmediately level 20.443

Expected level

expected level1.34confidence0.18

Between soon and immediately.The recorded answer, T = 1

The two levels are 0.011 apart, so it returns the expected level instead of pretending to know.

Returned"urgency": {"score": 1.34, "confidence": 0.18}0 output tokens

A routine: 0.104; B soon: 0.454; C immediately: 0.443.

Replayed from a response recorded on the MLX 8-bit build and published in the manchego-serve README; no model runs in your browser (the temperature control only re-weights the recorded numbers). The model reads the whole prompt once and returns a probability for every offered option; it never generates a token.
View as table
QuestionOptionRecorded probability
Which team? (choice)A = returns0.998457
B = shipping0.000967
C = billing0.000577
Is it defective? (noul)A = yes0.991509
B = no0.008491
How urgent? (score)A = routine0.103718
B = soon0.453758
C = immediately0.442524
JevBench
0.8053= of 28 open ~4B models
Calibration error
0.153 → 0.040on tasks it never saw
Smallest build
2.4 GBMLX 4-bit

EEG & brain–computer interfaces

Closer to the
neural signal.

Foundation models for EEG, packaged for the device that records it. Reconstruct missing channels with ZUNA, extract representations with CBraMod, and classify cue-paced motor imagery, all in Core ML, all checked against PyTorch.

Fig. 02 · Illustration

Fourteen electrodes, three missing

ZUNA takes the channels that were recorded, with their 3D positions, and estimates the ones that were not by iterative denoising.

5 cmTop view · t = 3.53 s

Denoising step 20/20
F3measured
F4estimated
T7estimated
P8measured
O2estimated

Seconds · synthetic, not recorded EEG

The head is the FreeSurfer fsaverage average scalp, seen from above and drawn to scale, with the real EPOC X electrode positions (MNE standard_1005). Traces and scalp field are synthetic, drawn to show the task: not recorded EEG and not model output. The field is a spherical-spline interpolation at one instant, t = 3.53 s, chosen for illustration. The green discs around T7, F4 and O2 are a drawing convention for emphasis, not the model’s region of influence.

02How we release

Every claim has a number. Every number has a boundary.

01Seal the test.v2.1’s recipe and its reading rule were frozen before a set of 25 unseen tasks was opened, once.02Gate the port.Core ML conversions are checked against PyTorch before they ship: 20-step diffusion parity for every ZUNA profile, 210 of 210 real trials for motor imagery.03Pin the bytes.manchego-serve checks the SHA-256 of every weight file and reports, in every response, whether it loaded the published ones.04Lead with the weaknesses.The Manchego card lists nine, and names the sets where v2.1 is worse than v2.

03Open models

Built here. Shared openly.

All 8 models
Manchegov2.1A 4B model for typed decisions. State, question, and options in; a probability for every option out, in one forward pass.Transformers · bf169.3 GBManchego · MLX 8-bitv2.1Manchego v2.1 for Apple silicon. 8-bit MLX, 4.5 GB, within one decision of the bf16 weights on the format test.MLX · 8-bit4.5 GBManchego · MLX 4-bitv2.1The smallest Manchego build: 2.4 GB for memory-constrained Apple devices, with its accuracy cost measured and stated.MLX · 4-bit2.4 GBManchego · LoRAv2.1The Manchego v2.1 adapter exactly as trained. Rank 32 on the last 16 layers of Qwen3.5-4B: 32.5M parameters, 130 MB.PEFT · LoRA adapter130 MBZUNA1.1 · Core MLfp32 profilesZyphra’s ZUNA1.1 on Apple platforms: EEG denoising, missing-channel reconstruction, and spatial superresolution in Core ML.Core ML · fp32380M parametersCBraMod · Motor imageryCore MLA cue-paced left/right motor-imagery classifier for a 14-channel EPOC X-style montage, in a single Core ML model.Core ML · fp3214 channelsCBraMod · EmbeddingsCore MLThe pretrained CBraMod backbone as a native Core ML embedder. A 14-channel, five-second window in; a 14,000-dimensional embedding out.Core ML · fp3214,000 dimensionsZUNA · Core MLFive profilesThe original ZUNA reconstruction model as five enumerated Core ML profiles, from a 14-channel EPOC X layout to 64 channels.Core ML · fp16380M parameters

Each model card documents provenance, evaluation, and limitations.8 models shown

04From the lab

Notes and signals.

All news
Open source

manchego-serve: Manchego, offline.

A 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.

Model update

Manchego, against the whole field.

The 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.

Model update

Manchego, on Apple silicon.

The 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.

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.

From the blog

Decision modelsDraftA decision is not a sentence.EEG & BCIDraftOne EEG model is not the whole interface.Lab notesDraftThe model is only half the release.

Follow the work

Good questions. Measured answers.