# Drex Max v1 (slim)

> Drex Max v1 is Nace.AI's open-weight 27B decision model, a fine-tune of Qwen3.8-27B that takes text, JSON and images with typed questions and returns a probability for every option without generating text.

- Full: https://www.anchorterminal.com/tools/drex-max-v1.md (~8,200 tokens) · this version ~2,430 tokens · JSON https://www.anchorterminal.com/tools/drex-max-v1.json · canonical https://www.anchorterminal.com/tools/drex-max-v1
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-30 · pricing https://www.anchorterminal.com/tools/drex-max-v1#pricing · reviews https://www.anchorterminal.com/tools/drex-max-v1#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 55.5/100 · rank #90 of 120 open-weight models · #17 in Decision models · not agent-ready · confidence low** · same model listed separately: Drex 1.5, Model API (https://www.anchorterminal.com/tools/nace-drex.min.md)


Assessment: A 27B cousin for typed decisions over text and screenshots, several questions per request with a probability for each option. Images need an NVIDIA GPU and Nace's loader; on Apple silicon only the text backbone runs, from 24 GB. The licence isn't open source, with free commercial use only under $1 million in revenue and funding.

## Facts

- Kind: Open-weight model · vendor: Nace AI, Inc. (Nace.AI) · category: Decision models · legal entity: Nace AI, Inc. · provenance 58/100
- Packages: huggingface `nace-ai/drex-max-v1`
- Auth: None · pricing: Free · x402: no · licence: Nace.AI Open RAIL-M License (modified), version 1.0, October 2026, in the repository's LICENSE. Not an open-source licence. Use is free only for individuals and organisations under $1 million in revenue in the prior 12 months and under $1 million in total funding (personal and research use excepted), never by anyone with a product that competes with Nace's, decision models included, and under use restrictions that bind the model, its derivatives and its outputs. Built on Qwen/Qwen3.8-27B (Apache-2.0), whose licence text isn't carried in the repository
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nace-ai/drex-max-v1 and the publisher's Q8_0 and affine 4-bit builds, read from the cards and their history, the files and their metadata, config.json, decision_config.json, the loader in code/drex_max, inference.py, the licence, Nace's GitHub docs and MLX script, Nace's blog, product page and hosted docs, the Jev Decision Index Space's data and Artificial Analysis's index on 30 October 2026. Nothing was downloaded or run. Nace's hosted API is the nace-drex listing
- Built on: Qwen/Qwen3.8-27B (Apache-2.0) at revision 1d4bf0f, named in decision_config.json and the loader, and in the card's metadata until 30 October 2026. A decision fine-tune of the backbone with a new 255-way readout head. Nace doesn't say how it was trained
- Relation to Drex 1.5 and the hosted API: The same publisher's larger, multimodal line. nace-drex is graded as Nace's hosted Drex API at console.nace.ai/v1/systemone, which today serves drex-max (this model, released 13 October 2026), drex-max-v1.1, drex-v2.0 (the default) and drex-v1.5, and forwards the retired drex-v1.0 to drex-max. Whether the hosted drex-max is byte for byte the Hub checkpoint isn't stated
- Model: 26,085,330,160 parameters in BF16 by the Hub's count (the card says 27B; the base counts 27.78B with its language-model head), 64 layers with 48 Gated DeltaNet layers that stay causal and 16 full-attention layers made non-causal, the Qwen3.5 vision tower, and a 255 by 5,120 readout over the last hidden state with a temperature of 1.009
- Files: 11 safetensors shards (52.17 GB), readout.safetensors, decision_config.json, config, tokenizer, processor config and the base's chat_template.jinja, the loader package code/drex_max, inference.py and requirements.txt. No pickle files. The affine 4-bit build adds pointer.bin, the readout in Nace's raw KPS1 format, not a pickle. The Hub lists a SHA-256 for every LFS file, and the builds' cards list theirs
- Licence: Nace.AI Open RAIL-M (modified) 1.0. Royalty-free copyright and patent grants, subject to Attachment A, which excludes users above $1 million in revenue or funding (personal and research use excepted) and anyone with a product competing with Nace's, decision models and document intelligence included, and bars fully automated decisions with legal or similar effects without human review, discrimination, unsupervised medical, legal or financial advice, military, weapons and mass-surveillance use, fraud and removing the licence. Derivatives, distillations and outputs carry the restrictions. No clause lets Nace change the terms. The Apache-2.0 base's licence text isn't in the repository
- Runtimes: Nace's loader on released Transformers 5.17.0 (pinned in requirements.txt) runs it as trained with images, on CUDA. The text backbone also runs on Nace's llama.cpp and Ollama forks (branch drex-max-v1), Kapitsa 0.2.25 or later (Nace's Apple silicon engine, Apache-2.0) and Nace's drex_max_mlx.py on MLX. No released llama.cpp, Ollama, vLLM, SGLang or LM Studio route, and no Ollama library entry
- Context: 262,144 tokens in config.json (rope default) and the docs, the base's native window. The loader defaults to 16,384 tokens per request (inference.py in the repository defaults to 8,192) and accepts up to 262,144 with DREX_MAX_CONTEXT or --max-length. The fine-tune's training length isn't stated
- Inputs and outputs: A state (text or JSON), optional images (file paths or data URLs, up to 4,194,304 pixels each), and named questions, each choice (1 to 255 options with descriptions), noul (yes or no, criteria optional) or score (ordered levels). Each question is scored as one pass over the state and returns probabilities, with choice and confidence, P(true), or an expected score and confidence. The system prompt tells the model to treat the state as data. No text generation and no thinking
- Hardware: The card's tested hardware is an NVIDIA H100 80 GB in BF16 (48.6 GiB loaded), an NVIDIA L40S 48 GB with the backbone in 8-bit (25.9 GiB), and an Apple M5 Pro with 48 GB for the Q8_0 GGUF on Metal. The affine 4-bit build is stated for Macs with 24 GB or more
- Training data: Not described. The card, the GitHub page and the blog say nothing about the decision fine-tune's data, size or cut-off. The nace-ai datasets for this model hold Decision Index results, not training data

- Scores: Capability pending, Reach 65, Risk 45, Agent interface 80, Documentation and upkeep 41 · negative events -2 · total over the 7 assessed categories
- Why:  · Reach, Memory 25 of 30, Nace's affine 4-bit build is 14.5 GB, 17.4 GB with a fifth added, so it fits 24 GB, though it reads text only. · Risk, File safety 15 of 25, safetensors and GGUF only, but the documented path runs the repository's loader code. · Agent interface, Read as a decision model, by what it takes and returns. · Documentation and upkeep, Model card 11 of 20, four of the seven items.
- Sources: 33, open questions: 6, both in the full twin
- Capabilities: inference.decision, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/drex-max-v1.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/drex-max-v1.svg` or a link to https://www.anchorterminal.com/tools/drex-max-v1 from a page on nace.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.
- Support: Discord https://discord.com/invite/nace-ai; https://trust.nace.ai; more in the full twin

## Before you call it

1. Clone github.com/nace-ai/drex-decision-models, download nace-ai/drex-max-v1 and run `python -m drex_max --model weights/drex-max-v1 --request examples/request.json` on an NVIDIA GPU
2. Plan for about 49 GB of GPU memory in BF16, or about 26 GB with --8bit on CUDA, before any long context
3. Don't load it with stock Transformers, vLLM or llama.cpp. Its full-attention layers must run non-causally with readout.safetensors applied, which only Nace's loader and forks do
4. On a Mac with 24 GB or more, run `kapitsa-systemone --model ./drex-max-v1-affine-q4 --port 8000 --context 2048` for text-only decisions at POST /v1/systemone
5. Raise the 16,384-token default with DREX_MAX_CONTEXT=N or --max-length N, up to 262,144, and expect more memory
6. Check the licence first. Above $1 million in revenue or funding, or with a competing product, you need Nace's commercial licence

## Connect

```bash
git clone https://github.com/nace-ai/drex-decision-models && cd drex-decision-models
hf download nace-ai/drex-max-v1 --local-dir weights/drex-max-v1
pip install -r models/drex-max-v1/requirements.txt
python -m drex_max --model weights/drex-max-v1 --request examples/request.json   # NVIDIA GPU, Python 3.12+; add --8bit for about 26 GB
```

```bash
kapitsa-systemone --model ./drex-max-v1-affine-q4 --port 8000 --context 2048   # Apple silicon, text only, after brew install nace-ai/tap/kapitsa and hf download nace-ai/drex-max-v1-affine-q4
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| JEV-27B-VL | C | 59.5 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/jev-27b-vl.min.md |
| joey-1 | C | 57.1 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/joey-1.min.md |
| mammoth-1 | D | 47.7 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/mammoth-1.min.md |
| elephant-1 | D | 47.6 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/elephant-1.min.md |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)