Decision model · v3

Manchego · MLX 8-bit

V3 on Apple silicon.

Manchego V3 for Apple silicon. 8-bit MLX, 4.5 GB, with the SemIf decision contract and published format checks.

Model: Manchego V3 · 4B
Weights: 8-bit · 4.5 GB
Runtime: MLX · Apple silicon
Trained options: 2–16 · SemIf
License: Apache-2.0
Release: 30 Sep 2026

Manchego V3 at 8 bits

This repository publishes the 4.5 GB 8-bit V3 conversion. The vision tower and multi-token-prediction head are removed; load the text-only weights with MLX LM.

Use SemIf, the contract V3 trained on. The old V2.1 short prompt is not an alternate way to obtain the V3 results.

Read the conversion’s own results

Published V3 format checks: SemIf, one question per forward pass, temperature 1. These small development checks do not measure general task transfer.

V3 format	Size	80 task rows · accuracy / CE	200 text rows · accuracy / CE
bf16 · Transformers	9.3 GB	0.963 / 0.141	0.785 / 0.602
MLX · 8-bit	4.5 GB	0.963 / 0.136	0.785 / 0.602
MLX · 4-bit	2.4 GB	0.887 / 0.239	0.765 / 0.715

The task rows come from families V3 trains on. The text check uses development rows from earlier corpora: three also appear in V3’s training data, and 50 come from banking77, which V3 did not train on. Read these as conversion checks, alongside the public and sealed evaluations on the main card.

Pin V3 and state its contract

Apple silicon · manchego-serve 0.2.0 with MLX installed

# In a manchego-serve v0.2.0 checkout, with the appropriate backend installed
manchego-serve-download --repo oraculumai/Manchego-MLX-8bit --revision v3 --out ./manchego-v3
manchego-serve --model ./manchego-v3 --revision v3 --contract semif --temperature-map off --backend mlx

The explicit --temperature-map off policy gives temperature-1 outputs. Inspect per-question temperature metadata instead of assuming the bf16 serving map applies to a different format. Score one prompt at a time when exact reproducibility matters.

The main model’s limits still apply

SemIf is the trained contract for 2–16 options and a nonempty state. Larger menus, up to 255 options, and empty states use a state-first fallback that V3 did not train on.
The longest training prompt was 6,386 tokens. Longer inputs are untested; the server’s larger input limit is not evidence of model quality at that length.
English text was trained and evaluated. Images and unrestricted text generation are outside the measured decision interface.
Use application-specific checks and human review for consequential decisions. Arithmetic and date calculations belong in code.

V3 remains overconfident on unfamiliar tasks and regresses on some computed-family draws. The main card describes training, evaluation, calibration and disclosures. Four-bit quantization adds a measured numerical trade-off.

Versioned conversion

Release revision: 778cc7b870ac74c2676efbf803c8baf28c13922e
Group size: 64
Licence: Apache-2.0
Model card reviewed: 2026-10-23

The original model and its complete release card remain the reference. This conversion is a separate numeric artifact, with its own published format card.

About this page

Summarized from the public model card and its published files. Tagged release 30 Sep 2026; reviewed 23 Oct 2026. The Hugging Face card is the source of truth for licenses and current details.

Upstream: Oraculum / Manchego V3