# Artificial Analysis: multimodal decision model Drex Max ranks #1. Both achieving SOTA and Pareto frontier. — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/f10cc103c53d6ce3

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-08 16:44 ingested 2026-11-09 02:37

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xr2rqk/artificial_analysis_multimodal_decision_model/)

## retained text



```
  
hey, its Alti from Nace.  
We released a small update with **Drex Max v1.1**, both achieving the SOTA and Pareto frontier for decision models. We did all that by iterating step by step and every previous version of Drex’s open weights are available on hf.  


https://preview.redd.it/8a4xlurt1f0i1.png?width=2000&format=png&auto=webp&s=99f721e992f967a1b263f9d7e1628ed70c20b1dd

we also released a full playbook on how we trained Drex Max in this paper: **“ Towards Foundation Decision Models: Drex”**

hf: [https://huggingface.co/papers/2611.07854](https://huggingface.co/papers/2611.07854)  
arXiv: [https://arxiv.org/pdf/2611.07854](https://arxiv.org/pdf/2611.07854)

Drex Max v1 is weights are available here: [https://huggingface.co/nace-ai/drex-max-v1](https://huggingface.co/nace-ai/drex-max-v1) with Open RAIL-M license & v1.1 open source weights upcoming soon (we are currently optimizing for local inference).

Three cool things about Drex Max models:

1. it can even run on your Mac with [https://github.com/nace-ai/kapitsa](https://github.com/nace-ai/kapitsa) inference engine.
2. Context window is \~256k, comparing to Glide1.2(64k) and Mammoth(16k) & pplxdecider’s (256k)
3. Its 8x cheaper than the next best model in the category in the aa index.

feel free to chime in: our chief of ai will answer any of your technical questions
```



document fetched 2026-11-09 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xr2rqk/artificial_analysis_multimodal_decision_model
