Nace posts Drex v1.5 weights. 58.08 is Nace’s; the license is not Apache

What did Nace post, and on which date?

The Hugging Face API object we opened for nace-ai/drex-v1.5 reports createdAt 28 September 2026 at 21:05 UTC, lastModified 10 October 2026 at 00:14 UTC, and 8,953,803,264 BF16 parameters. The card is public. It names the base XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, license other / nace-ai-open-rail-m, and a pointer head in head.pt.

The runtime repo is a later object. GitHub’s API for nace-ai/drex-decision-models reports created_at 9 October 2026 at 18:30 UTC, pushed_at 10 October at 00:14 UTC, and Apache-2.0 on the code. That repo is docs, runners, and model pages. The weights stay on Hugging Face.

OpenRouter’s model page for nace-ai/drex-v1.5, opened on 10 October, prints “Released Oct 9, 2026,” 131K context, and DeepInfra at $0.04 per million input tokens. That is OpenRouter’s listing date, not Nace’s hosted release. Nace’s catalog still says drex-v1.5 was released 2026-09-28 at $0.05 per million, with a 131,072-token state limit. We did not open a DeepInfra product page; that URL returned 404. For the hosted Jev that Drex’s card compares against, see TypeSafe’s $870 million Series A note. For Microsoft’s same-week Foundry scorer, see Microsoft-Decision-1.

How does Drex v1.5 score a request?

The card describes a decision model, not a chat model. It reads a state (text or JSON) and typed questions — choice, noul (yes/no), or ordinal score — and returns a probability for every option in one forward pass per question. “Nothing is generated.” Temperature, top_p and top_k in generation_config.json “do not apply.”

The backbone is listed as Qwen3_5ForCausalLM, 32 layers, hybrid attention (three linear-attention layers for every full-attention layer), about 9 billion parameters, about 18 GB on disk in BF16. Default context is 16,384 tokens, up to 131,072. The wire format is POST /v1/systemone, the same System One contract the card says the hosted Drex API and Drex DLM use.

Drex DLM is a different checkpoint. The runtime README lists it on Efficient-DLM-8B under CC BY-NC 4.0. Do not read the v1.5 RAIL-M card as covering DLM. A typed probability is also a different object from generated JSON; see structured AI output and Clef-omni, Cloudflare’s same-week omni decision head.

Which scores are Nace’s, and which board edition?

The card’s headline is “Drex v1.5 is the #1 model under 10B parameters on the public Decision Index 0.3.1 (58.08).” The table prints Drex 58.08 and Jev 1.13.0 at 57.96. Nace writes that the Drex row is “from our own run of the official kit on the full public suite” and that the other rows are the leaderboard’s. It also writes that Drex, Jev and Nimble “are within the board’s 0.9-point tie band,” and that Drex is ahead of Jev on 20 of the 37 benchmarks. Those sentences are Nace’s.

The living product page at nace.ai/drex, opened the same morning, still headlines Decision Index 0.2.1. It prints Drex 1.5 at 58.28 against Jev 1.13.0 at 57.91, and says Drex is ahead of Jev on 21 of the 38 benchmarks, with a 0.25-point tie rule on that edition. It also prints GPQA Diamond 45.4% against Jev’s 78.6%, and ACOS per-review F1 7.4% against 29.5%. Those weak-spot cells are on the 0.2.1 page. They are not reprinted as 0.3.1 numbers on the card we opened.

JevBench on both objects we compared is the same vendor pair: Drex 86.2% (199/231 on the product page) against Jev 87.0% (201/231). Long-document rows match across the card and the product page: 89.5% at 8k–32k tokens versus 76.5% truncated to 8k, and 93.4% at 32k–128k versus 78% truncated. The game-arena line is 122–47–87 (56.8%) against Jev. Every one of those figures is Nace’s. We did not rerun the kit.

What does Nace.AI Open RAIL-M actually block?

LICENSE on the card is “NACE AI OPEN RAIL-M LICENSE (MODIFIED), Version 1.0, October 2026,” issued by Nace.AI, Inc. “Open” in the preamble means the model is accessible, including commercial use by individuals and smaller organizations below Attachment A. It is not Apache. For why that split matters, stay on open weights versus open source and the model-license checklist.

Attachment A, paragraph 2, withholds a license if you, your employer, or an affiliated entity generated more than $1 million in gross revenue in the prior twelve months, or raised more than $1 million in equity or debt, except where use is limited to personal use or research. It also withholds a license if you provide a product or service that competes with Licensor’s, “including decision models and document intelligence services.” Commercial licenses go through nischay@nace.ai.

Use restrictions also bar fully automated decisions that produce legal or similarly significant effects on individuals without meaningful human review. Complementary Material — source code, scripts and documentation — “is not licensed under this License.” The card and runtime README put the Kev code under Apache-2.0. Do not flatten those two grants.

Where can you run it, and what is not a chat endpoint?

The card lists three supported runners: Python inference.py / serve.py on CUDA; a nace-ai/llama.cpp fork on the drex-v1.5 branch; and a nace-ai/ollama fork that needs that llama-server. serve.py listens on 127.0.0.1 and exposes POST /v1/systemone with no Authorization header. The card tells you to keep it there unless you put authentication in front of it.

Hosted Drex is a different path. The card says https://drex.nace.ai with bearer keys nace_sk_... and model drex-latest. The models page we opened says drex-latest currently points to drex-v1.5 and moved there on 28 September. A request billed under the alias is billed at the answering version’s price. Output tokens are not billed.

OpenRouter’s page is explicit that this model “runs on the OpenRouter Decisions API rather than the OpenAI-compatible chat endpoint” and that “chat completions SDKs will not work with it.” The listed input/output price there is $0.04 / $0 per million. That is $0.01 below Nace’s own $0.05 hosted row. We did not send a Decisions request, create a nace_sk_ key, or convert a GGUF.

What did we not test?

We did not download the safetensors, load head.pt, run inference.py, or start serve.py. We did not build the llama.cpp or Ollama forks, and we did not call drex.nace.ai, OpenRouter, or DeepInfra. We did not open the live multimodalart/jev-decision-index space as a finished public board. Secondary write-ups that mix Drex DLM with v1.5 are not additional evidence for this card.