Short answer
NACE AI's Drex is not one model but four: 8B (DLM), 9B (v1.5), 35B total with 3B active (v2.0), and 27B (Max). All four are open decision models, with weights and code published. Max takes the model ID drex-max.
They all do the same job. A request is a state (a document, record or agent context, optionally with images) plus a set of questions about it. Each question asks for a choice among options, a yes/no judgment, or a score on an ordered scale, and the model returns a probability distribution over the options with no token-by-token decoding. So this is not a comparison of who chats better — none of them chat.
Three things decide it:
- Images in the state → Max only. In the paper, only Max received vision tuning.
- The published benchmark numbers → Max only. The other three have no comparative figures at all.
- A decision model you can run yourself on limited hardware → start small. 8B is lightest; 35B and Max need real machines. The paper reports no accuracy for those three, so whether they are good enough is something only your own data can answer.
drex-max is not on TeamoRouter yet. Once it is live, the same key and model: drex-max is all you change; until then there is no same-vendor model on the site to use.
The whole comparison in one table
| Drex 8B (DLM) | Drex 9B (v1.5) | Drex 35B (v2.0) | Drex Max | |
|---|---|---|---|---|
| Parameters | 8B | 9B | 35B total / 3B active | 27B |
| Type | Open decision model | Open decision model | Open decision model | Open decision model |
| Vision tuning | No | No | No | Yes |
| Published benchmarks | None | None | None | Yes |
| Typical latency | — | — | — | 183 ms |
| Weights and code | Open | Open | Open | Open |
model value |
Not published | Not published | Not published | drex-max |
| On TeamoRouter | Not yet | Not yet | Not yet | Not yet |
What all four share:
- Input: one state (document, record or agent context, optionally with images) plus a set of questions.
- Question formats: pick one option / yes-no / score on an ordered scale.
- Output: a probability distribution over the options, a judgment, or a score — never free-form text.
- Training: supervised fine-tuning plus RLAF (reinforcement learning from adjusted feedback).
- Context window: no token count published for any of the four.
- API price: not published for any of the four.
Do not read the names as a ladder
This is the easiest mistake to make with this family.
8B is DLM, 9B is v1.5, 35B is v2.0, and 27B is Max. The version numbers and the parameter counts do not run in the same direction: v2.0 at 35B total is larger than Max at 27B. So do not assume "v2.0 is newer than v1.5, therefore stronger, and Max is the bigger v3." The release material sets out no such ordering.
The paper does exactly two comparative things:
- It lists the parameter count for all four sizes, and notes that Max alone received vision tuning.
- It publishes Max's benchmark results, calling it the strongest among the decision models compared. The other three have no numbers in the paper.
In other words, there is no official accuracy you can compare across the four. To find out whether 8B is enough, you have to run your own states and questions through it and look at where the answers land. That is not quite the usual "pick the cheap one, upgrade if it fails" reflex, because there is no ruler to measure against.
Why the price column is empty
There is no official API price to report, for any of the four. The release material has no per-token rate, no subscription, no free allowance and no hosted API — the models shipped as open weights, so the cost is the hardware and time it takes you to run inference.
Two consequences:
- The only published cost signal is parameter count. 8B is far lighter than the rest, 9B one step up, and both 35B and 27B need real inference hardware. The release gives no hardware requirements, so VRAM and throughput are yours to estimate from your own load.
- Ignore any "Drex Max costs $X per million tokens" figure. It does not generate output token by token, so that number cannot come from this release.
And when hosted pricing does exist, the right unit will not be per million tokens anyway — it will be per decision, because a Drex call returns one forward pass of probabilities and does not grow with answer length. Comparing it directly to a chat model's token rate will lead you to the wrong conclusion.
Choosing by task
Ask these in order; they are easier to answer than "which is best."
- Can you write the task as a choice? Classification, routing, ranking by priority, pass/fail, which of A or B is better — if it fits options, a yes/no, or a scale, Drex applies. Copywriting, code generation and open-ended Q&A are not that, so reach for a general-purpose model instead.
- Does the state contain images? With images, it has to be Max. This is the one hard dividing line in the family, because only Max was trained for vision.
- Do you need "good enough" or the published number? For the latter, only Max has a citable result. For the former, start at the small end and save on hardware.
- What can your machine run? 8B and 9B are the easiest to stand up; 35B and Max are a different order of magnitude.
A practical order: run a batch of your real states and questions through 8B and see whether the judgments hold up; if not, move to 9B; if the state carries images, start with Max. For the 35B with 3B active, the release only describes its parameters and says nothing about deployment differences, so how it behaves on your workload is also something you measure yourself.
Max against general-purpose models
The paper gives three sets of numbers for Max. Worth knowing, but not a cross-category ranking:
| Test | Max's result |
|---|---|
| Decision Index 0.4.1 public suite | 67.52, highest among the decision models compared (Jev 1.14.0 at 60.70) |
| Decision Bench v4 | Statistically tied with d3-turbo and GPT-6 Sol; 1.6 points behind Claude Opus 5.5 |
| DecideBench | Highest accuracy among decision models |
| Typical latency | 183 ms (GPT-6 Sol about 2.0 s, Claude Opus 5.5 about 3.5 s) |
Two things to keep in mind when reading them:
- These are the publisher's own tests. The numbers come from the paper, not from independent reproduction.
- They measure how well a model picks among the same options, not who writes better. A general model tying Max on Decision Bench means it can do that kind of task too; the difference is that Max answers with a single forward pass of probabilities — fixed latency, nothing to parse, and stable when you reorder the options (the paper adds consistency losses aimed exactly at that).
So the real question is not "is Max better than a general model." It is "is this judgment task worth a dedicated model." The answer usually comes down to latency and stability, not accuracy.
On TeamoRouter
You cannot call it yet: drex-max has not entered the catalog, and none of the four is on the site.
Once it is live, the key and base URL stay put and only model changes. How the state, questions and options are laid out in a request follows the interface docs after launch; Drex Max API shows where model goes and gives an example, so it is not repeated here.
drex-maxis not on TeamoRouter yet. Once it is live, the same key andmodel: drex-maxis all you change; until then there is no same-vendor model on the site to use.
How to pay
No subscription: top up your balance in the console from $5, billed by the tokens each request actually uses. Cards are supported, and so are USDT, USDC and ETH. The same key also calls Claude and GPT models.
New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.
FAQ
What actually separates the four?
Parameter count, plus the two things only Max has: vision tuning and published benchmarks. Input and output shape, training method and open weights are identical across all four.
Which is cheapest?
There is no official price, so there is no per-unit cost to compare. Self-hosted, cost tracks parameter count: 8B is lightest, then 9B, while 35B and Max need proper inference hardware.
Does a bigger version number mean stronger?
No. v2.0 is 35B total, larger than Max at 27B, and the naming encodes no strength ordering. The paper also publishes benchmarks for Max only.
How much better is 9B than 8B?
The release gives no comparative figures for either, and the parameters are 8B and 9B. You have to measure the gap yourself.
When do I have to use Max?
When the state contains images. Only Max received vision tuning.
Can any of them write code or articles?
No. All four answer over the options, yes/no questions or ordered scales you supply, and none of them produces free-form text.
How long is the context window?
No token count is published for any of the four. The release only says a request carries a state and a set of questions.
Which one can I use today?
None — no Drex model has entered TeamoRouter's catalog. Once drex-max is listed, the same key and a new model ID is all it takes.
Related reading
- Drex Max is out: the model ID, what was actually announced, and how to integrate
- Using Drex Max without an official account: what you can do with no account and no foreign card
- The paper: the official release material