Drex Max Finance

Drex Max Finance is the finance version of Drex Max, from Nace.AI. It is a 27B multimodal decision model for banking, accounting, payments, insurance, credit, compliance, filings and markets. You give it a state (text, JSON, or images such as invoices, receipts and statements) and typed questions, and it returns a calibrated probability for every option in a single forward pass. It generates no text, and it never answers with an option you did not supply.

• 77.3 on Nace's Finance Index (48 held-out public finance test sets), 11.5 points above Drex Max v1.1 and 18.9 above GPT-6 Luna; 85.7% of 3,361 finance decisions right.
• 27B dense, fine-tuned from Drex Max v1.1 (built on Qwen3.8-27B) with non-causal full attention and a 255-way readout head.
• Context up to 262,144 tokens. States longer than 8,192 tokens are cut down to the passages most relevant to your questions before the model reads them, and you are billed only for the tokens it reads.
• TypeSafe System One compatible: the same request and response as Drex Max and Jev.
• Multimodal: text, JSON and images. Every question on a request sees all of its images.

What it is for

Area	Example decisions
Banking & payments	intent and routing for customer messages, card / transfer / dispute triage, fraud and phishing flags
Accounting & audit	invoice and receipt fields, expense policy checks, account classification, reconciliation exceptions
Credit & insurance	applicant risk bands, default flags, claim routing, policy questions
Compliance	KYC / AML red flags, sanctions and disclosure checks, ESG and climate claims
Filings & reports	10-K / XBRL tagging, numeric claims, table and text QA over financial reports
Markets	news and social sentiment, entity sentiment, monetary-policy stance, M&A and event signals

Question types

Type	You send	You get back
noul	a yes/no question, with optional criteria.true / criteria.false	noul, the probability of "yes"
choice	options as criteria (name → description or null), 1-255 options	choice (the most likely option), probabilities per option, and confidence
score	ordered levels as a criteria list, 1-255 levels	score (the expected level), probabilities per level, legend, and confidence

One request can ask up to 64 questions about the same state.

Example

curl https://api.deepinfra.com/v1/decisions \
  -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nace-ai/drex-max-finance",
    "state": {
      "transaction": {"amount": 9800, "currency": "USD", "type": "cash_deposit", "branch": "Miami-04"},
      "history": "Customer made 4 cash deposits between $9,500 and $9,900 in the last 10 days."
    },
    "questions": {
      "aml": {"type": "choice", "instructions": "Which AML typology does this activity best match?",
              "criteria": {"structuring": "Deposits kept just under the reporting threshold",
                           "funnel": "Many depositors, one account", "none": null}},
      "sar": {"type": "noul", "instructions": "Should this be escalated for a suspicious activity report?"},
      "risk": {"type": "score", "instructions": "Customer risk rating", "criteria": ["low", "medium", "high"]}
    }
  }'
copy

The response holds answers keyed by question name and usage (input_tokens, output_tokens).

Images

Add images, a list of up to 4 images, each either a base64 data URL ("data:image/jpeg;base64,...") or an object {"content_type": "image/jpeg", "base64": "..."}, in a request body of up to 13 MiB; remote image URLs are not fetched. Larger images are scaled down to about 1.6 MP.

IMG=$(base64 -w0 invoice.jpg)
curl https://api.deepinfra.com/v1/decisions \
  -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nace-ai/drex-max-finance",
    "state": {"po": "PO 7731: 50 pallets at USD 964 each", "receiving": "41 pallets logged"},
    "images": ["data:image/jpeg;base64,'"$IMG"'"],
    "questions": {
      "action": {"type": "choice", "instructions": "What should AP do with this invoice?",
                 "criteria": {"pay": "Pay in full now", "partial": "Pay only for the pallets received",
                              "hold": "Hold payment pending review"}},
      "mismatch": {"type": "noul", "instructions": "Does the invoice disagree with receiving?"}
    }
  }'
copy

The same API is served at https://api.deepinfra.com/v1/systemone.

Performance

Figures reported by Nace.AI.

Finance held-out suite

48 public finance test sets, 3,361 decisions, none seen in training. Every set has equal weight; scores are chance-corrected skill × 100, and unanswered items count as wrong.

Benchmark	Drex Max Finance	Drex Max v1.1	Drex Max	GPT-6 Luna
Finance Index	77.3	65.8	60.1	58.4
Finance accuracy	85.7%	77.7%	73.1%	70.9%
Sentiment & tone	80.7	69.3	58.6	68.6
News & headlines	89.5	71.4	68.3	60.7
Banking & support intent	87.1	74.5	65.6	71.9
Documents & extraction	92.3	87.8	67.2	48.8
Financial QA & filings	96.7	89.7	84.5	88.1
Credit & risk	48.3	0.0	21.7	38.3
Monetary policy	53.8	30.0	25.0	38.8
ESG & climate	73.0	59.4	60.0	48.1

GPT-6 Luna ran at low reasoning effort with text input only, so image-based items count as unanswered.

General decision quality

Benchmark	Drex Max Finance	Drex Max v1.1
JevBench (231 items)	94.4%	93.9%
JevBench, hard tier	89.2%	87.4%
Decision Bench (1,071 cases)	94.6%	94.2%

License and use restrictions

Drex Max Finance is © Nace.AI, Inc. Its weights are under the Nace AI OpenRAIL-M license, which permits non-commercial research and internal evaluation; other uses of the weights need a license from Nace.AI. DeepInfra serves this model in partnership with Nace.AI.

Its answers are decision support. Keep meaningful human review on credit, insurance, compliance and other consequential financial decisions. Drex and Nace are trademarks of Nace.AI, Inc.