Drex Max v1: Nace's Free 27B Decision Model, Now on Your Mac

Your AI agent does not need another essay. It needs a yes or a no.

Every support ticket, fraud check, and moderation flag is a decision, and most of the industry answers those decisions with a chat model that writes a paragraph, names an option somewhere inside it, and hopes a parser can find it. Nace AI has been arguing for a different shape: a model that takes a case and typed questions, and returns a probability for every option in a single forward pass. Nothing is generated.

On November 1, 2026, the company released the open weights of Drex Max v1, its most capable decision model: a dense 27-billion-parameter network that reads text, JSON, and images, and serves the same decision API whether it runs on an NVIDIA H100 or a 24GB Mac. The weights are free to download, with a license catch worth reading before you build a business on them. Here is what the release actually contains, what the numbers say, and what it costs to run.

What a decision model is, and why your agent might want one

Drex Max does not chat. You send a state, the facts of a case in text, JSON, or images, plus one or more named questions, and it returns probabilities. Questions come in three types: a choice among options you name, a yes/no, and an ordinal score on a scale you define. Several questions about one case can share a single pass.

Nace's own documentation puts the difference plainly: an LLM writes text token by token and names an answer at the end, while Drex returns probabilities with no token stream, so there is no JSON to repair, no truncation, and no invented options. It can only answer with the options you gave it.

The canonical example comes from the official model card: a support ticket reading "I was charged twice for the same order. Please refund the extra payment." goes in; the model returns billing at 0.9988, technical at 0.0003, other at 0.0009, plus a refund yes/no and an urgency score, all from one request. If you run routing, triage, guardrails, or agent tool selection, that is the shape of output you actually consume downstream.

What landed on November 1

The release ships the model in several builds. The full bf16 weights run about 49GB on disk. A Q8_0 GGUF build comes in at 27.25GB. An affine 4-bit checkpoint targets Apple Silicon. Alongside the weights, Nace published forks of llama.cpp and Ollama on a drex-max-v1 branch, plus support in Kapitsa (v0.2.25 or later) and MLX, all documented in the official Drex Max v1 model card. All five runtimes serve the same POST /v1/systemone API.

The release caps a fast six weeks for the Drex line. Nace launched Drex 1.0 on September 24, followed it with Drex 1.5 on September 28 (open weights arrived by early October), and put the Drex Max API up in mid-October. The November 1 weights release is the first time the flagship model's weights are public. One related launch not to confuse it with: Nace's October 26 "Drex for Search" announcement is a separate product for agentic search, not this weights release.

One precision note, because Nace's own site can confuse this: nace.ai now leads with Drex Max v1.1, a newer hosted revision scoring 70.8 on the Decision Index. The open weights released November 1 are v1, which scores 67.52. Everything below about downloadable weights refers to v1; hosted figures and v1.1 scores are labeled as such.

Under the hood: 64 layers, no tokens

The backbone is a dense 27B network with 64 layers: 48 Gated DeltaNet linear-attention layers that run left to right, and 16 full-attention layers, one every fourth layer, that are non-causal, so every token can attend to the whole sequence. Images pass through a native vision tower and are visible to every question in a request. A 255-way readout head turns the last hidden state into probabilities over up to 255 options per question.

The config lists a 262,144-position window, though the default per-request limit is 16,384 tokens; raising it takes an environment variable (DREX_MAX_CONTEXT=N) or a flag. One warning straight from the model card: the bundled loader applies the non-causal attention mask, and a plain causal chat run will not reproduce the published scores. If you wire this into llama.cpp yourself, use the fork.

The numbers — all reported by Nace

Every benchmark figure in this section comes from Nace's own materials, and is flagged as such. The headline: 67.52 on the public Decision Index 0.4, against 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0. The public index averages 37 benchmarks across five areas; Nace says all 140,178 scoreable requests were answered, and unanswered requests count as wrong. The area breakdown shows where the model is strong and where it is not: Tools 81.9, Language 72.2, Retrieval 68.8, Knowledge and Reasoning 57.1, Arts 53.4.

On Artificial Analysis's Decision Index, Drex Max scores 64, fourth of 18, one point behind the three leaders at 65, and ahead of Claude Haiku 5.5 (Low) at 61 and GPT-6 Luna (Low) at 55. It also posts 92.2% on JevBench and 94.30% on Decision Bench v4.

Vision is a genuine strength: a 73.92 public vision index across 11 benchmarks and 12,210 rows, with 97.3 on InfographicVQA and 92.4 on document extraction (KIE). The weak spots are on the card too: 40.6 on Hateful Memes moderation, 38.6 on R-Bench-M, 47.3 on MMMU-Pro vision. This is not the model you pick for hard multimodal reasoning or moderation alone.

One partial independent cross-check exists. The community-run Jev Decision Index board (edition 0.4.1, data from October 22) lists Drex Max v1 (27B) with a full score of 66.6, ranked 15th of 229, a public-suite score of 67.52 matching Nace's headline, and a calibration error of 0.057. Nace's number checks out against the public suite; the full private-suite score is a touch lower, which is normal for vendor headlines.

The license catch: read this before you download

Drex Max v1 is open-weight, not open source — and the distinction is doing real work here. The weights ship under the Nace AI Open RAIL-M License (Modified), version 1.0, dated October 2026. I read the full text. The key terms:

• Commercial use is free only below two thresholds: your organization generated $1 million or less in gross revenue in the prior 12 months, and has raised $1 million or less in total funding. Above either line, commercial use is not licensed; you need a separate written license from Nace (contact listed as [email protected]).
• You may not use the model in any product or service that competes with Nace's offerings, including decision models and document intelligence services — regardless of your revenue.
• Fully automated decisions with legal or similarly significant effects (employment, credit, housing, insurance, healthcare, education, immigration, law enforcement) require meaningful human review. No military, weapons, or mass surveillance uses; no fraud.
• Derivatives must be distributed under the same license, and the use restrictions travel with the model, its derivatives, and its outputs.

For a solo developer or a small team under the thresholds, the license is permissive, including commercial use. For a funded startup or an enterprise, the weights are a free evaluation, not a free lunch. Read Attachment A of the LICENSE file yourself before shipping to customers.

What it really costs

The weights cost $0. The bill is hardware.

Full bf16 (~49GB on disk) needs serious iron: Nace's tested setup is an NVIDIA H100 80GB, with 48.6 GiB loaded. Drop to 8-bit and you need about 26GB, tested on an NVIDIA L40S 48GB at 25.9 GiB loaded, with probabilities within 0.03 of bf16 on the company's example.

On a Mac, the math is friendlier. The affine Q4 checkpoint runs on Apple Silicon Macs with 24GB of memory or more via Kapitsa or MLX. The Q8_0 GGUF needs 48GB, through the llama.cpp fork, the Ollama fork, or Kapitsa. Nace's own test: on a 48GB Apple M5 Pro, the Q8_0 GGUF answered within 0.02 of the H100 bf16 run per probability, and the affine Q4 build stayed within 0.0329 of Q8_0 with no choice flips. All self-reported, but the direction is clear: local hardware can approximate datacenter output for this workload.

If you would rather not self-host, Nace's own Drex page lists Drex Max v1.1 at $0.03 per 1,000 decisions, at a 50% launch price, so expect $0.06 when the launch period ends. Drex 2.0 sits at $0.06 and Drex 1.5 at $0.04 per 1,000 decisions on the same page. For comparison, the page lists Jev 1.14 at $0.042 and GPT-6 Luna Decisions at $0.10 per 1,000. Note the units: decision models bill per decision, not per token, which is why the numbers look tiny next to token-priced APIs. For a sense of the decision-model beat, we recently covered Milliseconds.ai's $0.04 decision API and OpenAI's Decisions API going GA; both bill differently and neither ships open weights. One more fine-print item: on Mac, CPU, and GGUF paths, only text is supported — image inputs need the Python path on an NVIDIA GPU.

Setup: three paths, from the official docs

These steps are condensed from Nace's model card — the card's own phrasing, shortened. Nothing here is tested hands-on; treat it as a guided tour of the official instructions.

Python + NVIDIA GPU. Install the Hugging Face CLI, download the weights (hf download nace-ai/drex-max-v1), install requirements, and run python -m drex_max --model weights/drex-max-v1 --request examples/request.json. Add --8bit for the ~26GB path. Needs Python 3.12 or newer; images work here.

Mac, the easy path. Run brew install nace-ai/tap/kapitsa, download the affine Q4 repo (hf download nace-ai/drex-max-v1-affine-q4), then kapitsa-systemone --model ./drex-max-v1-affine-q4 --port 8000 --context 2048. Query it with a POST to localhost:8000/v1/systemone. Needs 24GB+ of unified memory; text only.

Ollama fork. Clone Nace's ollama and llama.cpp forks on the drex-max-v1 branch, build llama-server, convert or download the Q8_0 GGUF, and write a Modelfile with a CAPABILITY decision line. Set num_gpu 40 to offload 40 layers to Metal on a 48GB Mac, or num_gpu 0 for CPU.

Drex Max v1 vs the field

*Decision-1 bills per input token, not per decision — the units do not convert cleanly. All figures are Nace-reported except the Artificial Analysis index and the community-board cross-check. Microsoft's rate is from our October 25 decision-API coverage.

Where it does not fit

• Images on local paths. Mac, CPU, and GGUF builds are text only for now; vision needs the Python path on an NVIDIA GPU.
• Moderation and hard reasoning. The card's own weak spots: 40.6 on Hateful Memes, 38.6 on R-Bench-M, 47.3 on MMMU-Pro vision.
• The license. Funded companies and anyone building a competing product need a commercial deal with Nace.
• No independent head-to-heads yet. Everything above is Nace-run, with one community-board cross-check as partial corroboration.
• v1.1 already exists on the hosted side. The weights are v1. If you need the newest scores, the hosted API is currently the only route.

Pro tips and common mistakes

• Do not run the weights as a causal chat model and compare scores: the card warns the bundled non-causal loader is required to reproduce results.
• The 262,144-token window is not on by default: requests cap at 16,384 unless you set DREX_MAX_CONTEXT.
• Do not download the 27GB GGUF expecting vision — images are not in the file. Only the Python path sees images.
• On a 48GB Mac, Nace recommends --context 2048 for Kapitsa; larger contexts eat the memory the model needs.
• Read Attachment A of the LICENSE before you ship: the $1M revenue/funding thresholds and the no-competing-products clause are the two lines most likely to surprise a team.

Frequently asked questions

Is Drex Max v1 really open source?

No. The weights are public under Nace's Open RAIL-M v1.0 license, which permits commercial use below $1 million in revenue or funding, bars competing products, and carries use restrictions that travel with derivatives and outputs. That is open-weight, not OSI open source.

Can I run Drex Max v1 on my Mac?

Yes, for text decisions. The affine Q4 checkpoint needs 24GB or more of unified memory (Kapitsa or MLX); the Q8_0 GGUF needs 48GB. Image inputs require the Python path on an NVIDIA GPU.

How much does Drex Max cost?

The weights are free to download. Nace's hosted API lists Drex Max v1.1 at $0.03 per 1,000 decisions at a 50% launch price, so $0.06 after launch. Self-hosting costs whatever the GPU or Mac costs you. There is no per-token billing, because the model bills per decision.

How is Drex Max different from OpenAI's Decisions API?

OpenAI's Decisions API is a configured LLM endpoint billed per token ($0.10 per million input tokens, output free). Drex Max is a dedicated decision model: a 27B backbone with a 255-way readout head that returns probabilities in one forward pass, with open weights you can run yourself. One is an API feature; the other is a model you can own.

What is the difference between Drex Max v1 and v1.1?

v1 is the open-weights release from November 1, scoring 67.52 on Decision Index 0.4. v1.1 is a newer hosted-only revision on nace.ai scoring 70.8. Nace has not released v1.1 weights.

Sources

• Nace AI — Drex Max v1 model card (official; benchmarks, architecture, hardware, setup)
• Nace AI Open RAIL-M License v1.0, October 2026 (official; read in full for the license section)
• Nace.AI — Drex product page (official; v1.1 figures, hosted pricing, latency)
• Nace.AI homepage (official; Drex for Search announcement, October 26)
• MarkTechPost — "Nace AI Open-Sources Drex Max v1" (November 1; disclosed as sponsored by Nace AI)
• Wisevoter — "Nace AI Open-Sources 27B Drex Max Decision Model" (November 1)
• SystemOneModels — Drex family reference (Jev Decision Index board data, release timeline)
• AI Core Updates — Claude Haiku 5.5 (referenced in the AA index comparison)