Nace AI Drex: Open-Weight Decision Models, and the Licence Catch

Nace AI released Drex Max v1, a dense 27B open-weight "decision model" that returns a calibrated probability for every answer option in a single forward pass instead of generating text, under the Nace.AI Open RAIL-M licence rather than an open-source licence. That licence bars use by any entity above $1 million in revenue or funding raised, except for personal or research use, and bans competitors — a limit the post says matters more than any benchmark. Hosted access on OpenRouter costs $0.06 per 1M input tokens and $0 per 1M output tokens, and a Q4 build targets 24 GB Macs, a Q8 GGUF 48 GB, and bf16 H100-class GPUs.

Nace AI has released Drex Max v1 as open weights: a dense 27B model that never writes a sentence. You give it a situation and a list of typed questions, and it returns a probability for every answer option in one forward pass. The weights are public on Hugging Face, but the licence is Nace.AI Open RAIL-M , not an open-source licence. The licence file bars use by any entity above $1 million in revenue or funding raised, apart from personal or research use. That limit matters more than any benchmark number below. This post separates three things: what the sources confirm, what Nace reports about itself, and what you should check before you build on the model. If you are new to the category, read our guide to what decision models are https://explainx.ai/blog/what-are-decision-models-ai-classifiers-guide-2026 first. Quick answers | Question | Answer | |---|---| | What is it? | A dense 27B "decision model": state plus typed questions in, per-option probabilities out. | | Who made it? | Nace AI. Drex Max succeeds Drex v2.0. | | Is it open? | Weights are public. The Open RAIL-M licence carries use limits, including a $1M revenue and funding ceiling personal and research use excepted and a ban on competitors. | | Can I run it locally? | Yes. A Q4 build is aimed at 24 GB Macs, a Q8 GGUF at 48 GB, and bf16 at H100-class GPUs. | | How much does hosted access cost? | $0.06 per 1M input tokens, $0 per 1M output tokens on OpenRouter. | | Are the scores independent? | No. Nace reports them. Artificial Analysis lists one score 64 , which Nace cites. | What does a decision model actually do? A rail switch sending one track to one of three destinations, like a Nace AI Drex decision model routing a request A normal LLM answers by generating tokens. A decision model answers by scoring. According to the OpenRouter listing https://openrouter.ai/nace-ai/drex-max?ref=explainx , Drex Max takes a state in text, JSON, or up to four images, plus typed questions. The questions can be yes/no, multiple choice, or an ordinal score. The output is "a calibrated probability for every option in a single forward pass, with no text generation." The listing says it is compatible with the TypeSafe System One interface and suits "routing, classification, and policy checks that need a scored answer instead of prose." This design has practical consequences: - Speed and cost. One forward pass, no decoding loop. That is why the hosted price has an output rate of $0. - Repeatability. You read a probability, not a paragraph you must parse. You can set thresholds, such as "escalate if P fraud is above 0.3." - Fit for guardrails. A policy check "does this request break rule 4?" is a yes/no question with a number attached. - No explanation. You cannot ask it why. Pair it with a normal LLM when a human needs a reason. Several vendors now ship this pattern. We covered Liquid AI's LFM3 family https://explainx.ai/blog/liquid-ai-lfm3-open-weight-2b-family-text-vision-audio-omni-decision-models-2026 and Cloudflare Clef-Omni https://explainx.ai/blog/cloudflare-clef-omni-open-multimodal-decision-model-2026 in October, and the Amazon Bedrock AgentCore guide https://explainx.ai/blog/amazon-bedrock-agentcore-evaluations-decision-models-jev-strands-decider-screening-layer-guide-2026 shows how a decision model can act as a screening layer in an agent pipeline. Drex is the same idea from another lab, at a larger size. What is in the Drex Max v1 release? The Hugging Face model card https://huggingface.co/nace-ai/drex-max-v1?ref=explainx is the primary source. It is Nace's own page, so treat its framing as vendor framing. It states: - A dense 27B backbone with 64 layers. Forty-eight are Gated DeltaNet linear-attention layers, and 16 are full-attention layers placed every fourth layer. The full-attention layers are non-causal. - A native vision tower. Images stay visible to every question in a request. - A 255-way readout head, so one question can have up to 255 options. - A qwen3 5 model type and a 262,144-position window, with 16,384 as the default. - Files: bf16 weights, a Q8 0 GGUF, an affine Q4 checkpoint, and forks of llama.cpp and Ollama. Local Mac and CPU paths are text only for now, according to the model card: the GGUF has no images, and llama.cpp needs the Python path for images. If you need image questions, plan on an NVIDIA GPU path. Which files fit which machine? A laptop with a small chip, showing a 24 GB Apple Silicon Mac running a quantized Drex Max decision model locally | Build | Size per the model card | Target hardware | |---|---|---| | bf16 | about 49 GB | NVIDIA H100 80 GB tested | | 8-bit | about 26 GB | NVIDIA L40S 48 GB tested | | Q8 0 GGUF | 27.25 GB | Mac with 48 GB | | Affine Q4 | not stated on the card | Mac with 24 GB or more | Nace documents five runtimes: Python, Kapitsa, MLX, llama.cpp and Ollama. The local servers expose a POST /v1/systemone API. On a 48 GB Mac the company recommends --context 2048 for its Kapitsa runtime. For Ollama, set num gpu 40 to offload 40 layers to Metal, or num gpu 0 to run on the CPU. The Ollama fork also needs a CAPABILITY decision line in the Modelfile. If local inference on a Mac is new to you, our explainer on llama.cpp https://explainx.ai/blog/what-is-llama-cpp-run-models-locally-2026 covers the runtime family these forks come from. For hardware context, see the Apple M6 local AI specs https://explainx.ai/blog/apple-m6-imac-macbook-pro-14-local-ai-specs-prices-november-6-2026 post. How good is it? Scores, and who reported them Every number below is self-reported by Nace unless stated. We did not find an independent replication. | Benchmark | Reported score | Notes | |---|---|---| | Decision Index 0.4 public text | 67.52 | Drex v2.0: 61.25. Jev 1.13.0: 57.96. | | Decision Index sub-scores | Tools 81.9, Language 72.2, Retrieval 68.8, Knowledge and Reasoning 57.1, Arts 53.4 | From the model card. | | Public vision index | 73.92 | 11 benchmarks, 12,210 rows. InfographicVQA listed at 97.3. | | Artificial Analysis Decision Index | 64 | Nace says this ranks fourth of 18. | | JevBench | 92.2% 213/231 | Self-reported. | | Decision Bench v4 | 94.30% 1010/1071 | Self-reported. | One detail deserves attention. The results cover 140,620 requests, and unanswered requests count as wrong. That is a stricter rule than some benchmarks use, which is good. But the numbers still come from the vendor's own harness. The score scales also differ. The Nace figure is on its own public Decision Index 0.4, and the Artificial Analysis Decision Index is a separate average of AA-Omniscience-Decision, HLE-MultiChoice, When2Call and WildGuardTest. Do not compare numbers across indexes or versions. Our earlier posts on benchmarks that shift under you https://explainx.ai/blog/artificial-analysis-intelligence-index-v5-opus-5-5-58-private-benchmarks-open-weight-gap-12-points-2026 make the same point for general LLM indexes. Does quantization change the answer? For a decision model this is the key question, because you pick the highest-probability option. According to the model card, on a 48 GB Apple M5 Pro the Q8 0 GGUF answers the example request within 0.02 of the bf16 weights on an NVIDIA H100, and the 8-bit load on an L40S is within 0.03 on the same example. That is one example request, not a benchmark, and it comes from Nace's own tests. The card gives no such figure for the Q4 build. Run your own labelled sample before you trust it on your task. The licence catch: Open RAIL-M is not open source This is the part most headlines drop. The LICENSE file on Hugging Face https://huggingface.co/nace-ai/drex-max-v1/blob/main/LICENSE?ref=explainx is titled "Nace AI Open RAIL-M License Modified ", version 1.0, October 2026. Attachment A bars use for any purpose by a person or entity that: - generated more than $1,000,000 in gross revenue in the prior twelve months, - has raised more than $1,000,000 in total equity or debt funding, or - offers a product or service that competes with Nace's, "including decision models and document intelligence services". The first two limits carry an exception for personal use or research purposes. The licence also bans fully automated decisions with legal or similarly significant effects such as hiring, credit or housing without meaningful human review, and bars military, weapons and mass-surveillance use. Section 4 says use above the thresholds needs a separate written commercial licence from Nace. What that means in practice: - Hobbyists and researchers: personal and research use is carved out of the revenue and funding limits. - Small startups under both thresholds: commercial use is allowed, unless you compete with Nace's own products. - Anyone above the thresholds, or a competitor: you need a separate commercial licence. The licence file points to nace.ai/drex for this. This is our reading of a legal text, not legal advice. If you plan to deploy Drex Max, have the licence reviewed. Compare this with Cloudflare Clef-Omni https://explainx.ai/blog/cloudflare-clef-omni-open-multimodal-decision-model-2026 , which is Apache-2.0, and with open agent models like ByteDance's Axion 1.0 https://explainx.ai/blog/axion-1-0-bytedance-open-agent-model-two-sft-stages-rl-18k-examples-35b-122b-apache-2026 . Those licences let a large company ship the weights. Drex Max does not, on the reported terms. Where does Drex Max sit in the Drex line? The model card links two earlier releases, Drex v2.0 and Drex v1.5. OpenRouter https://openrouter.ai/nace-ai/drex-max?ref=explainx describes Drex v2.0 as a 35B mixture-of-experts decision model with 3B active parameters, and Drex v1.5 as a 9B model. It calls Drex Max the largest decision model from Nace.AI and the successor to Drex v2.0. Drex Max is a dense 27B model, so "largest" refers to its place in the lineup and its score, not to the biggest total parameter count. We saw other reports of further Drex models and training claims, but we could not confirm them on a Nace page, so we leave them out. What people are asking Can I use it as an agent router? That is the main use case Nace and the listings describe. A router asks one multiple-choice question "which tool fits this request?" and reads the top option. Because the answer is a probability, you can add a fallback: if the top option is under a threshold, send the request to a bigger LLM. Our agent tool selection coverage https://explainx.ai/blog/amazon-bedrock-agentcore-evaluations-decision-models-jev-strands-decider-screening-layer-guide-2026 shows the same screening pattern. How does it compare with a small LLM prompted to classify? A small LLM can classify too, but it generates text, it can drift in format, and it can flip its answer when you reorder options. A decision model fixes the output format by design. It does not fix bias or accuracy, and order sensitivity can still exist. You still need your own evaluation set. Does it hold up on long context? The model card, as summarized, lists a 262,144-position window with 16,384 as the default. Linear-attention layers help with cost, but long-context state tracking is hard for many models. See our post on AI2's StateBench https://explainx.ai/blog/ai2-statebench-long-context-state-tracking-gated-deltanet-frontier-models-13-percent-2026 , which tested Gated DeltaNet style models on that exact weakness. We have no Drex-specific long-context test, so assume nothing past the default window until you measure. What does it cost to try? Locally, the Q4 build runs on a Mac with 24 GB or more. Hosted, OpenRouter https://openrouter.ai/nace-ai/drex-max?ref=explainx lists $0.06 per million input tokens and $0 per million output tokens, a 262K context, and an October 21, 2026 release date. The listing shows one provider, DeepInfra, with a 50% discount on its posted rates. The Hugging Face repository was created on October 13, 2026, so the listing date is the hosted-API date, not necessarily the first weights upload. A practical test plan before you adopt it 1. Build a labelled set. Collect 200 to 500 real examples of the decision you want to automate, with the answer a human gave. 2. Run Q4 and Q8 side by side. Check that choices match, not just probabilities. 3. Shuffle option order. Count how often the top choice changes. 4. Check calibration. If the model says 0.8, is it right about 80% of the time? Plot it in ten buckets. 5. Read the licence. Confirm your company is under the revenue and funding limits, or plan to talk to Nace. 6. Compare with an Apache-licensed option. Cloudflare Clef and Liquid LFM3 are alternatives with different terms. What this means for what you build or pay If you run an agent that spends LLM tokens on routing and triage, a decision model can cut both cost and latency. The hosted price of $0.06 per million input tokens with free output is small next to a frontier LLM call. The catch is the licence and the self-reported benchmarks. For a small team under the thresholds, Drex Max is worth a weekend test. For a larger company, start with the Apache-licensed options, and treat Drex as something to revisit if Nace publishes commercial terms. The wider trend is clear: vendors are racing to make "scored answers" a standard tool beside chat models. Expect more releases, and expect the Decision Index to be argued over. Independent testing, such as Artificial Analysis listings, will matter more than launch-day tables. Related reading Details in this post are accurate as of November 6, 2026. Licence terms, model sizes and scores may change; check the official model page before you deploy.