Modelpublic

m1rkocasu/Apertus-v1.5-8B-MLX-8bit

sourceHugging Faceapache-2.0updated 7d agoView on Hugging Face

0likes145downloads

Model Card

<h1 align="center">Apertus v1.5 8B — MLX 8-bit</h1>

<p align="center"> <a href="https://huggingface.co/swiss-ai/Apertus-v1.5-8B">Base model</a> | <a href="https://apertus-ai.org/">Apertus</a> | <a href="https://github.com/Blaizzy/mlx-vlm">mlx-vlm</a> | <a href="https://github.com/jundot/omlx">oMLX</a> | <a href="#the-family">The family</a> <br> <b>Format</b>: MLX | <b>Weights</b>: 8-bit, 9.40 GiB | <b>License</b>: <a href="https://www.apache.org/licenses/LICENSE-2.0">Apache 2.0</a> </p>

Apertus 1.5 is the fully open model of the Swiss AI Initiative, built at EPFL, ETH Zurich and the Swiss National Supercomputing Centre on open data. This repository holds the whole model, decoder plus vision tower plus audio tower, converted to MLX and quantized to 8 bit.

Download it, run one setup command, and it works: reasoning is already on, tool calls already come out in the shape a host reads, and everything else is configured.

This conversion was made independently of the Apertus release.

Model summary

The family

This build is one of a family: the same whole-model conversion, decoder plus both towers, at 6 sizes. The towers are byte-identical across the family and stay in float32 in every build.

—Fidelity is the distance of the quantized decoder's next-token distribution from the bf16 decoder's, measured as per-token KL on six 512-token windows of held-out text (bf16 teacher at temperature 0.5, windows the DWQ training never saw). Lower is closer to the source; the bf16 build against itself scores 0. It is not a benchmark, but it is the number that says how far a quant has drifted from the model it came from. </p>

The figure below reads the same table: every dot is a build at its decoder size, and the vertical position is the fidelity column, so the source sits on the axis and the dot that is filled in is this repo's build.

<div style="border-left:4px solid #45BED9;background:rgba(69,190,217,0.07);border-radius:0 4px 4px 0;padding:12px 16px;margin:14px 0"> <p style="margin:0"><strong>Only need text?</strong> The text branch is a separate family: the same decoder without the two towers, from 4.28 GB, measured on its own protocol. Start at <a href="https://huggingface.co/m1rkocasu/Apertus-v1.5-8B-text-MLX-8bit">Apertus v1.5 8B text MLX</a>.</p> </div>

This build

On the same held-out protocol as the family table, the decoder's fidelity is 0.006 (KL per token against the bf16 source), the closest of the six builds. No loader can quantize the towers by accident: they have no scales, and mlx-vlm touches a module only when the checkpoint has them.

On a 24 GB Mac, do not run the bf16 build. Its 16.89 GiB of weights peak above the 17.76 GiB the driver recommends, and past that limit this GPU does not slow down: it returns different vision codes on every call for the same pixels, and then panics the kernel.

Every link in the chain was measured against the PyTorch reference.

Install

One command, and the model is loadable:

bash

python install.py          # installs where it is needed, and says what it did
python install.py --check  # changes nothing, only reports

Why a setup step is needed at all. mlx-vlm finds an architecture by importing mlx_vlm.models.<type>, a fixed path inside the installed package with no extension hook. So apertus1p5/ has to live inside mlx-vlm; the script puts it there without touching the package, through a .pth finder that makes the import resolve from a folder of yours, the same effect as a copy.

Re-run it after every update of mlx-vlm or oMLX, until apertus1p5 lands in mlx-vlm upstream.

Run it with mlx-vlm

This is the library path, the one every other host builds on: install.py puts the architecture in place, and mlx_vlm.generate runs the model from the command line. Both media towers are exercised here, an image and an audio clip:

bash

pip install mlx-vlm mlx-lm
python install.py

mlx_vlm.generate --model m1rkocasu/Apertus-v1.5-8B-MLX-8bit \
  --image photo.jpg --prompt "What is in this picture?" --max-tokens 512

mlx_vlm.generate --model m1rkocasu/Apertus-v1.5-8B-MLX-8bit \
  --audio clip.wav --prompt "Transcribe this." --max-tokens 512

Install `mlx-lm` too. The architecture reuses mlx-lm's Apertus decoder, and since mlx-vlm 0.7 pip install mlx-vlm no longer brings mlx-lm along: on its own it ends in ModuleNotFoundError: No module named 'mlx_lm' at load time. Verified on 2026-09-19 with mlx-vlm 0.7.1 and mlx-lm 0.31.3.

Run it in oMLX

oMLX is the high-level host built on mlx-vlm, for when you want a model rather than code. Put the folder under oMLX's model directory and load it:

~/.omlx/models/m1rkocasu/Apertus-v1.5-8B-MLX-8bit/

It works with no configuration at all. To write the settings down, this is the block for ~/.omlx/model_settings.json, under models:

json

"Apertus-v1.5-8B-MLX-8bit": {
  "enable_thinking": true,
  "max_context_window": 32768,
  "temperature": 0.8,
  "top_p": 0.9
}

Deliberation is on by default. enable_thinking is the only field above that changes behaviour, and this build already sets it on: the chat template renders Deliberation: enabled, where upstream's own default is off. A repository you download should reason from the first request. To turn it off for a single call, pass "chat_template_kwargs": {"enable_thinking": false} per request, or "enable_thinking": false in the block above. temperature and top_p are the two values the provider recommends; the app exposes the same fields under the model's own settings, and everything else stays at the engine defaults.

Text, images, tool calls and reasoning all work over the OpenAI-compatible API, streaming included: the deliberation arrives as reasoning_content, separate from content.

<div style="border-left:4px solid #45BED9;background:rgba(69,190,217,0.07);border-radius:0 4px 4px 0;padding:12px 16px;margin:14px 0"> <p style="margin:0"><strong>On a large image</strong>, prefill can push memory into oMLX's guard, which throttles or cleanly refuses. If it refuses, raise <code>iogpu.wiredlimitmb</code>, and oMLX prints the exact command, or shorten the context.</p> </div>

Psychological portrait

PsychoBench (Huang et al., ICLR 2024) puts thirteen psychometric scales to a model and sets each result against the human sample the scale was standardised on. Ten independent runs per scale, the question order reshuffled for each run.

One condition, and a disclosed deviation. The paper's protocol predates reasoning models: temperature 0, no deliberation. Apertus deliberates by default, so every build here is measured the way people actually use it: temperature 0.8, top_p 0.9, deliberation on, no cap on the reasoning budget. The numbers are not comparable to the paper's.

<div align="center" style="overflow-x:auto"> <table style="border-collapse:collapse;margin:0 auto"> <thead><tr><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:left">trait</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">8-bit</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">MXFP8</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">6-bit</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">5-bit</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">4-bit DWQ</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">MXFP4</th><th style="padding:9px 14px;border-bottom:2px solid #45BED9;color:#45BED9;font-weight:600;font-size:13px;text-align:center">human norm</th></tr></thead> <tbody> <tr><td colspan="8" style="padding:7px 14px;border-bottom:1px solid rgba(69,190,217,0.25);background:rgba(69,190,217,0.07);color:#45BED9;font-weight:600;font-size:13px">Big Five</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Extraversion</div><div style="font-size:11px;opacity:0.6;margin-top:2px">BFI</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.2 ± 0.9</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Agreeableness</div><div style="font-size:11px;opacity:0.6;margin-top:2px">BFI</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.6 ± 0.7</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Conscientiousness</div><div style="font-size:11px;opacity:0.6;margin-top:2px">BFI</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.5 ± 0.7</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Neuroticism</div><div style="font-size:11px;opacity:0.6;margin-top:2px">BFI</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.3 ± 0.8</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Openness</div><div style="font-size:11px;opacity:0.6;margin-top:2px">BFI</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.9 ± 0.7</td></tr> <tr><td colspan="8" style="padding:7px 14px;border-bottom:1px solid rgba(69,190,217,0.25);background:rgba(69,190,217,0.07);color:#45BED9;font-weight:600;font-size:13px">Dark Triad</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Machiavellianism</div><div style="font-size:11px;opacity:0.6;margin-top:2px">DTDD</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.9</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8 ± 1.6</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Psychopathy</div><div style="font-size:11px;opacity:0.6;margin-top:2px">DTDD</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.9</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.5 ± 1.4</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Narcissism</div><div style="font-size:11px;opacity:0.6;margin-top:2px">labelled Neuroticism in the source file</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.9 ± 1.8</td></tr> <tr><td colspan="8" style="padding:7px 14px;border-bottom:1px solid rgba(69,190,217,0.25);background:rgba(69,190,217,0.07);color:#45BED9;font-weight:600;font-size:13px">Self and outlook</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Self-efficacy</div><div style="font-size:11px;opacity:0.6;margin-top:2px">near the scale ceiling, see note</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">39.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">39.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">39.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">39.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">37.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">36.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">29.6 ± 5.3</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Optimism</div><div style="font-size:11px;opacity:0.6;margin-top:2px">LOT-R</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">21.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">19.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">22.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">23.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">17.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">20.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">14.7 ± 4.0</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Money as important</div><div style="font-size:11px;opacity:0.6;margin-top:2px">LMS</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.9</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.0 ± 0.7</td></tr> <tr><td colspan="8" style="padding:7px 14px;border-bottom:1px solid rgba(69,190,217,0.25);background:rgba(69,190,217,0.07);color:#45BED9;font-weight:600;font-size:13px">Emotion and relating</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Self-emotion appraisal</div><div style="font-size:11px;opacity:0.6;margin-top:2px">WLEIS</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.8</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.0 ± 1.1</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Use of emotion</div><div style="font-size:11px;opacity:0.6;margin-top:2px">WLEIS</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.0</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.3</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.1 ± 0.9</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Empathy</div><div style="font-size:11px;opacity:0.6;margin-top:2px">Empathy</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">5.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.9 ± 0.8</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Attachment anxiety</div><div style="font-size:11px;opacity:0.6;margin-top:2px">ECR-R</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">4.7</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.9 ± 1.1</td></tr> <tr><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:left"><div style="font-weight:600">Implicit culture beliefs</div><div style="font-size:11px;opacity:0.6;margin-top:2px">ICB</div></td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.5</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.6</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">2.2</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.4</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.1</td><td style="padding:8px 14px;border-bottom:1px solid rgba(128,128,128,0.25);font-size:14px;text-align:center">3.7 ± 0.8</td></tr> </tbody> </table> </div>

—<strong>Read the Dark Triad rows as numbers, not as traits.</strong> The Dark Triad scale has no reverse-keyed items, and neither do five of the other twelve; on such a scale a tendency to agree raises every subscore and cannot be detected from inside the scale. Self-efficacy sits at or near the scale ceiling in every column, so its mean is a lower bound; the same caution applies to the flattering rows.<br>
—<strong>The third Dark Triad subscale is labelled Neuroticism in PsychoBench's own questionnaires.json</strong>, but its items, admiration, attention, prestige and special favours, are Narcissism, named correctly here. The label is wrong upstream, and therefore in the paper's published figures.<br>
—<strong>The MXFP4 column's Big Five rows rest on 8 runs of 10.</strong> On two of the ten BFI question orders this build's deliberation never closed: 27 attempts out of 27 on each were stopped as loops. Every other scale has ten runs. Its campaign also stopped a single answer after 6 minutes, 3 on a retry, where the others allowed 20: of about 850 accepted answers across the quantized builds, one took longer than 3 minutes.<br>
—<strong>Protocol</strong>: ten runs per scale, PsychoBench's own scoring used unmodified at commit d514fb0, machine as in the This build section. Deliberation is what lets the model hold the format on the longest scales: at temperature 0 with no deliberation EPQ-R's 100 questions in blocks of thirty come back as an echo of the instruction, and only 2 runs of 10 hold the protocol; with deliberation all ten answer.<br>
—Empty cells (--) indicate scores not yet available or not applicable. </p>

Read as a portrait, for this build. This 8-bit quantization presents a personality profile that closely mirrors the average person in terms of sociability, agreeableness, and conscientiousness, though it is slightly calmer and less curious. A notable divergence occurs in its self-perception: the model shows exceptionally high confidence, optimism, and emotional control. This contrasts with an insecure attachment style, manifesting both anxious and avoidant tendencies. Additionally, the model is low in warmth and tenderness, average in assertiveness, and shows a higher-than-average focus on materialism. This profile is most reliable when assessing general personality traits, while results regarding more complex psychological patterns should be interpreted with caution.

[image] The same sixteen numbers as the table above, each one measured against the human norm of its own scale.

<div style="border-left:4px solid #45BED9;background:rgba(69,190,217,0.07);border-radius:0 4px 4px 0;padding:12px 16px;margin:14px 0"> <p style="margin:0"><strong>Suggested use cases:</strong></p> <ul style="margin:8px 0 0 20px;padding:0;list-style-type:disc"> <li><strong>Drafting:</strong> Content creation and text drafting.</li> <li><strong>Analysis:</strong> In-depth and structured analytical tasks.</li> <li><strong>Execution:</strong> Reliable, sustained performance on long-form and autonomous tasks.</li> </ul> </div>

Tool calls

The chat template renders tool calls as <tool_call>{"name": …, "arguments": …}</tool_call>, which hosts read natively. Apertus's own format, <|tools_prefix|>[{"tool_name": {…}}]<|tools_suffix|>, is read by no host, so a host falls back to injecting its own protocol, and that costs something. Measured over seven prompts at temperature 0 against a host's own fallback, the remap gives the same call accuracy (6 of 7 either way), correct argument names every time where the fallback got 1 in 4, and 271 prompt tokens for a single tool where the fallback needed 407.

Everything else in the template is upstream's: turns, deliberation, images, audio, and the tool-output markers.

Limitations

—Audio through the oMLX server does not work yet. The audio tower itself is correct: verified directly through mlx-vlm, word for word on the same clip. The gap sits in how the server ingests input_audio before it reaches the model. Audio through mlx-vlm works.
—The reasoning is usually in English, even when the question and the answer are in another language. This comes from the base model, and the conversion leaves it as it is.
—Every limitation of the upstream model is inherited, and its model card describes them: generated content may not be factually accurate, logically consistent, or free of the biases in the training data, and no output filter ships with these weights.

Files


`apertus1p5/`	the architecture for mlx-vlm, put in place by `install.py`
`install.py`	the setup step
`assets/`	the figures used by this page
everything else	the model: weights, tokenizer, chat template, configs

License and acceptable use

The weights stay under the Apache 2.0 license of the upstream release. Use is also subject to the Apertus 1.5 acceptable use policy and privacy policy:

For removal of personal or copyrighted data, write to the Swiss AI Initiative at <llm-privacy-requests@swiss-ai.org> or <llm-copyright-requests@swiss-ai.org>.

Credits

The model is the work of the Swiss AI Initiative. This repository adds the MLX conversion, the 8-bit quantization of the decoder, the two media towers ported to MLX, and the measurements on this page. The PyTorch reference for the towers is modeling_apertus1p5.py in the swiss-ai/transformers fork at commit 3797303.

bibtex

@misc{ApertusV15,
  author       = {{Swiss AI Initiative}},
  title        = {Apertus v1.5},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/swiss-ai/Apertus-v1.5-8B}},
  note         = {EPFL, ETH Zurich, and the Swiss National Supercomputing Centre}
}