---
title: "Drex Max, the decision model that sees"
description: "Drex Max decides from screens, images and video, built for agents that use computers and browsers. It scores 67.59 on the public Decision Index, 6.9 points over Jev 1.14, decides in under 200 ms and accepts every screenshot size. Drex 2.0 ships open weights, and Drex 1.5 routes text at volume."
canonical: "https://nace.ai/drex"
---

# Drex Max, the decision model that sees

> Drex Max decides from screens, images and video, built for agents that use computers and browsers. It scores 67.59 on the public Decision Index, 6.9 points over Jev 1.14, decides in under 200 ms and accepts every screenshot size. Drex 2.0 ships open weights, and Drex 1.5 routes text at volume.

## Drex Max

The frontier Drex, built for agents that use computers and browsers. It reads screenshots, charts, forms, scanned documents, photos and video frames, not only text, and returns the same typed answers as every Drex model. API only: Console (https://console.nace.ai) and OpenRouter (https://openrouter.ai/nace-ai). No open weights.

- Model id: `drex-max` (pinned; `drex-latest` does not point to it). Send images in an `images` array; up to 8 per request.
- Context: 131,072 tokens per question, image tokens included.
- Price: $0.06 per million input tokens; output is not billed.
- Guide: https://console.nace.ai/docs/drex/guides/drex-max

Decision benchmarks, beside the other Drex models, Jev 1.14 and GPT-6 Luna Decisions. GPT-6 Sol and Claude Opus 5.5 are general-purpose frontier models, shown for reference; they are not scored on the Decision Index.

| Benchmark | Drex Max | Drex 2.0 | Drex 1.5 | Jev 1.14 | GPT-6 Luna Decisions | GPT-6 Sol | Claude Opus 5.5 |
| --- | --: | --: | --: | --: | --: | --: | --: |
| Decision Index 0.3.2, public (37 benchmarks, 140,178 scored requests) | 67.6 | 61.3 | 58.1 | 60.7 | 47.8 | – | – |
| JevBench | 92.2 | 85.7 | 84.8 | 86.7 | 91.0 | 99.1 | 100.0 |
| Decision Bench | 94.3 | 88.3 | 86.8 | 92.3 | 90.5 | 94.7 | 95.6 |
| DecideBench (item) | 98.5 | 96.5 | 90.0 | 98.0 | 97.2 | 99.2 | 98.8 |
| DecideBench (pair) | 97.0 | 93.5 | 80.5 | 96.0 | 94.5 | 98.5 | 98.5 |
| Game Bench (win %) | 64.0 | 60.3 | 58.7 | 50.0 | 57.5 | 65.0 | 63.0 |
| MMLU-Pro | 72.5 | 72.8 | 58.6 | 82.1 | 58.1 | 93.0 | 94.0 |
| HLE | 19.0 | 12.8 | 11.4 | 19.4 | 12.3 | 39.0 | 56.0 |
| CRUXEval | 85.8 | 84.6 | 76.7 | 74.7 | 66.4 | 100.0 | 95.0 |
| ARC-Challenge | 97.4 | 96.8 | 95.9 | 97.6 | 96.9 | 99.0 | 98.0 |
| ARC-Easy | 99.4 | 98.6 | 97.9 | 99.3 | 99.0 | 100.0 | 100.0 |
| HellaSwag | 97.9 | 95.9 | 92.1 | 94.3 | 91.4 | 93.0 | 96.0 |
| OSWorld-Verified (automation) | 0.780 | 0.698 | 0.676 | 0.734 | 0.747 | 0.756 | 0.790 |
| OSWorld-V2 (automation) | 0.737 | 0.636 | 0.640 | 0.687 | 0.659 | 0.744 | 0.761 |
| BFCL (tool calling) | 98.2 | 96.6 | 94.9 | 96.4 | 94.2 | 94.7 | 97.3 |
| ToolRet (tool retrieval) | 70.6 | 67.3 | 70.2 | 65.0 | 60.4 | 72.7 | 74.8 |
| When2Call | 85.2 | 86.5 | 89.7 | 81.7 | 70.2 | 78.3 | 86.7 |
| RAGTruth | 84.8 | 82.5 | 82.7 | 75.1 | 77.0 | 72.5 | 69.8 |
| Median API latency | 183 ms | 151 ms | 145 ms | 196 ms | 326 ms | ~2.0 s | ~3.5 s |
| Cost per 1,000 decisions | $0.06 | $0.06 | $0.04 | $0.042 | $0.10 | $3.33 | $10.17 |

Latency is the median API time per decision, as the caller sees it, network included; ~ marks an approximate figure.

Vision, on the public Decision Index Vision suite, against GPT-6 Luna Decisions:

| Benchmark                     | Drex Max | GPT-6 Luna Decisions |
| ----------------------------- | -------: | -------------------: |
| Decision Index Vision, public |     74.5 |                 63.9 |
| CV-Bench                      |     86.3 |                 73.8 |
| BLINK                         |     74.1 |                 63.7 |
| RealWorldQA                   |     84.0 |                 75.0 |
| Winoground                    |     93.0 |                 83.8 |
| CharXiv                       |     86.1 |                 78.8 |
| InfographicVQA                |     98.8 |                 93.8 |
| KIE (CORD + FUNSD)            |     95.5 |                 97.5 |
| Mind2Web                      |     88.6 |                 81.2 |
| R-Bench-M                     |     48.9 |                 32.5 |
| MMMU-Pro (vision)             |     53.6 |                 41.2 |
| Hateful Memes (moderation)    |     70.6 |                 71.2 |

Business decisions, against Jev 1.14: FinEntity 92.3 vs 81.9, BANKING77 90.0 vs 79.7, ContractNLI 75.9 vs 57.7. Jev 1.14's API rejects images with a side under 280 px or over 1,568 px; Drex Max accepts every screenshot. Median API latency per screenshot decision: 174 ms against 253 ms for Jev 1.14, one request at a time, on the same 243 screenshot questions from the same client.

## The family

| Model | Model id | Best at | Where |
| --- | --- | --- | --- |
| Drex Max | `drex-max` | Decisions from screens, images and video; the most accurate | Console and OpenRouter |
| Drex 2.0 | `drex-v2.0` | Open weights: run it on your own hardware | Console, or self-hosted from GitHub and Hugging Face |
| Drex 1.5 | `drex-v1.5` | Lean and quick, high-volume text routing | Console |

The sections below cover Drex 2.0 and earlier unless they name Drex Max.

## Trusted by

Intel, General Catalyst, Barton, Big 4, Prosper, Mountain America Credit Union.

## Get started

- Base URL: `https://console.nace.ai`
- Auth: `Authorization: Bearer $NACE_API_KEY`
- Model: `drex-latest`
- Open models: https://github.com/nace-ai/drex-decision-models (runtime, docs and weights to run Drex yourself)
- Endpoint: `POST /v1/systemone` with `{"model":"drex-latest","state":"...","questions":{...}}`; `GET /v1/models` lists checkpoints.

```bash
curl https://console.nace.ai/v1/systemone \
  -H "Authorization: Bearer $NACE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"drex-latest","state":"Help! My payouts have been failing for 3 days.","questions":{"is_urgent":{"type":"noul","instructions":"Does this convey urgency?"}}}'
```

The API is wire-compatible with TypeSafe's Jev: keep `@typesafe-ai/sdk` and set `TYPESAFE_BASE_URL=https://console.nace.ai`, `TYPESAFE_API_KEY`, `TYPESAFE_DEFAULT_MODEL=drex-latest`.

## What it is

Drex is a decision model, not a chatbot. You send the facts of a case plus typed questions; one forward pass returns a probability for every option you supplied. There is no token stream, no chain of thought and no free-form JSON, so it can only ever answer with options you gave it.

Three question types:

| Type | You send | You get back |
| --- | --- | --- |
| `choice` | named options as a `{name: description}` map | `choice`, `confidence`, `probabilities{name: p}` |
| `noul` | a yes/no question | `noul` = P(true) in [0,1] |
| `score` | an ordered list of levels | `score` = expected level, plus a distribution over levels |

Several questions about one state share a single pass over that state.

## How it differs from a chat model

In the language of dual-process psychology, a chat model is System 2: it writes its way to a decision one token at a time, then names an answer. A decision model is System 1: it has seen the shape of the decision before and answers in one pass, with a probability for each option.

|  | A chat model | Drex |
| --- | --- | --- |
| How it answers | Generates text, then names an answer at the end | Reads the case once and scores every option in the same pass |
| What comes back | A sentence, or JSON to parse | A probability for each option offered; nothing else can come back |

## How it differs from Jev

Jev is the decision model the public Decision Index is named after. Drex answers the same requests in the same shape. Measured against Jev 1.13.0:

- Decision Index 0.3.1, public benchmarks: Drex 2.0 at 61.25 vs 57.96 (+3.29), ahead in four of the five areas.
- JevBench (231 public items): Drex 2.0 198 correct (85.7%), Jev 200 (86.6%). The other board models were not run on it.
- Board games: Drex 2.0 won 132 to 76, with 48 draws, over 256 games (see Board games below).
- Tokens read per decision: Drex 1.1 median 70 vs 367 over eight identical playground requests (Jev read 2026-09-24, Drex 1.1 2026-09-26).

## Decision Index

The Decision Index is a public leaderboard of benchmarks that score decisions rather than prose (https://huggingface.co/spaces/multimodalart/jev-decision-index). Drex 2.0 and Drex 1.5 were scored by us with the official 0.3.1 kit on the full public suite: 37 benchmarks in five areas. The index is chance-corrected skill, so random guessing scores 0. These are public benchmarks only; the board's Full score also blends in private tests, which are not part of these runs.

The table shows Drex beside Jev and the strongest models under 10B, not the top of the board: eight dense models of 26B to 33B parameters score higher on the public index. Other rows are the board's published scores. Area columns are chance-corrected skill × 100.

| Model | Parameters | Public index | Raw | Knowledge & Reasoning | Language Understanding | Retrieval & Classification | Tools & Automation | Arts & Human Taste |
| --- | --: | --: | --: | --: | --: | --: | --: | --: |
| Drex 2.0 | 35B, about 3B active | 61.25 | 70.55 | 47.9 | 65.3 | 63.0 | 76.5 | 53.9 |
| Drex 1.5 | 9B | 58.08 | 68.08 | 44.6 | 60.4 | 61.7 | 75.0 | 48.8 |
| Jev 1.13.0 | undisclosed | 57.96 | 68.55 | 53.9 | 59.2 | 55.4 | 75.1 | 39.1 |
| Bespoke Nimble 9B v3 | 9.7B | 57.19 | 68.38 | 41.0 | 60.7 | 55.5 | 84.2 | 44.0 |
| Cloudflare clef-flash | 9.7B | 56.15 | 66.53 | 53.0 | 47.2 | 52.2 | 81.9 | 48.4 |
| ezjev 4B s2 | 4.7B | 50.82 | 63.16 | 31.0 | 60.5 | 56.3 | 69.9 | 31.1 |

Drex 2.0 is built on Qwen3.6-35B-A3B, a mixture of experts with 35B parameters in total and about 3B active per token. Drex is trained on the official training splits of the index benchmarks alongside verifiable procedural data, and evaluated only on the held-out evaluation splits.

## Long context

Drex 2.0 reads up to 128K tokens in one request, like Drex 1.5 and up from 32K on Drex 1.1. Measured on Drex 1.5: long documents with typed questions, every request answered, one forward pass per question:

| Context length | Accuracy, whole document | Accuracy, first 8K only | Median latency |
| --- | --: | --: | --: |
| 8K to 32K tokens | 89.5% | 76.5% | 0.65 s |
| 32K to 128K tokens | 93.4% | 78.0% | 2.0 s |

## Board games

A head-to-head we ran between Drex 2.0 and Jev 1.13.0 with the Kaggle Game Arena harness, OpenSpiel rules (not a Kaggle leaderboard result). Every legal move is offered as an option and each move is one pass over the board, with no look-ahead. Each game has 16 openings, each played from both sides, for 32 games per title and 256 in total. Drex 2.0 won 132, Jev 76, and 48 were drawn: an overall score of 60.9% with a draw counting half, up from 56.8% for Drex 1.5. Drex 2.0 won more games than Jev in six of the eight titles. The last column is Drex 1.5's score against Jev under the same protocol.

| Game | Drex 2.0 wins | Draws | Jev wins | Drex 2.0 score | Drex 1.5 score |
| --- | --: | --: | --: | --: | --: |
| Lines of Action | 21 | 11 | 0 | 82.8% | 85.9% |
| Chess | 13 | 19 | 0 | 70.3% | 60.9% |
| Checkers | 20 | 5 | 7 | 70.3% | 70.3% |
| Nine Men's Morris | 15 | 13 | 4 | 67.2% | 45.3% |
| Clobber | 20 | 0 | 12 | 62.5% | 53.1% |
| Reversi | 17 | 0 | 15 | 53.1% | 48.4% |
| Dots and Boxes | 15 | 0 | 17 | 46.9% | 56.3% |
| Four in a Row | 11 | 0 | 21 | 34.4% | 34.4% |

## Throughput

One H100 replica sustains ~25 req/s on a mixed corpus (no errors to 64 in flight); the fleet scales from 1 to 10 replicas.

## Demonstrations

- Doom deathmatch: Drex picks target, fire, movement and weapon every tick. 727 consecutive decisions, 0 errors.
- Lemmings: on the Steel Works level Drex assigns one skill at a time (blocker, builder and the rest) to a chosen lemming, opening a route for 72 lemmings to reach the exit.

## Deployment

Managed on our GPUs (shared or dedicated), inside your own AWS or Azure account, or air-gapped in your datacenter, with an identical request contract everywhere, so the modes can be mixed. It fits on a single accelerator. The Drex 2.0 open weights take about 70 GB in bf16 and about 37 GB as a Q8_0 GGUF for llama.cpp and Ollama (https://huggingface.co/nace-ai/drex-v2.0). Drex 1.1 also shipped as FP8 weights: on 2,511 questions FP8 gave the same answer as bf16 on 98.1%, moved probabilities by 0.015 on average (0.29 at most), and scored 92.57% accuracy against 92.49% in bf16. It can be tuned on your own labelled decisions and the weights shipped to you.

## Compliance

SOC 2 Type II audited. The report covers how Drex is built, deployed and supported, with its security controls tested over a period of operation. A penetration test report is also on the trust center (https://trust.nace.ai). Reports are available on request: https://nace.ai/contact?product=drex&source=drex-compliance

## Links

- Get an API key: https://console.nace.ai
- Product page: https://nace.ai/drex
- Contact: https://nace.ai/contact
- Community (Discord): https://discord.com/invite/nace-ai

## For agents

- [Drex API guide](https://nace.ai/drex.md): the decision model: request shape, question types, benchmarks and deployment, as markdown
- [Decision endpoint](https://console.nace.ai/v1/systemone): POST with model "drex-latest", a state string and typed questions (choice, noul, score); https://console.nace.ai/v1/models lists checkpoints
- [Document endpoints](https://console.nace.ai/v1/documents/{parse|split|classify|extract|ground}): POST a source (a public URL, an uploaded workspace file or a finished parse job) and poll GET https://console.nace.ai/v1/documents/jobs/{id}; the same key and credit as decisions
- [Console](https://console.nace.ai): API keys, the playground, usage and billing; send `Authorization: Bearer $NACE_API_KEY` on every request
- [Console docs](https://console.nace.ai/docs): quickstart, question types, the document routes, errors, limits and pricing
- [Nace Document Intelligence](https://nace.ai/ndi.md): the Perception series: parse, split, classify, extract and ground any file type with a location on every value
- [Developers](https://nace.ai/developers.md): The execution layer for agents: Perception, Decision and Context behind one key, the Drex Max frontier, install steps, a scripted agent session and the four deployment boundaries
- [Site index](https://nace.ai/llms.txt): every page on nace.ai; append `.md` to any page URL or send `Accept: text/markdown` for the markdown version