Not tested by us yet. Ginza 1.0 is in limited early access behind a waitlist, and we have no access, so our harness cannot call it. Everything below comes from De¹'s Financial World Model press release and reprints of its launch notice at PANews, KuCoin and TokenPost. We'll update this post if we get access or a public API. Nothing here is financial advice.
Ginza 1.0 is a model from De¹ that estimates, before you sign an on-chain transaction, how that transaction is likely to execute. It is the first model in De¹'s Financial World Model series. Access is limited to partners, invited users and vetted waitlist applicants, and De¹ has published no price, accuracy figure or technical detail. Right now it is a launch to watch, not a model to build on.
| Ginza 1.0 | |
|---|---|
| Maker | De¹ (announced from the British Virgin Islands) |
| Announced | 2026-11-05 |
| Type | Financial decision model (forecast of execution outcome) |
| Series | First model of the Financial World Model |
| Input | On-chain data, per the launch coverage |
| Access | Limited early access, waitlist on De¹'s website |
| Price | Not published |
| Size, architecture, benchmarks | Not published |
What it is
De¹ calls Ginza a decision model for on-chain finance. In the launch coverage the pitch is simple: instead of acting on a quoted price and a trader's instinct, you see a probability estimate of how the trade will actually play out, then decide whether to sign. Coverage uses both "probabilistic" and "deterministic" to describe the output. Those words conflict, and De¹ has not said which one it means.
The parent project is the Financial World Model, which De¹ announced on 3 August 2026. De¹ says it trains on real capital movements rather than synthetic data, and that it feeds signals from "hundreds of millions" of dollars in daily trading volume back into the system. It is meant as a general base that can be specialized for high-frequency trading, perpetual futures, prediction markets, tokenized real-world assets and other markets. Ginza 1.0 is the first public piece.
Access and who gets in
Applicants join a waitlist on De¹'s site. The first batch goes to De¹ Lab Alliance partners and invited users. Other applicants are reviewed on trading quality, sustained activity and long-term contributions to the platform. That is an unusual gate for a model: it selects users by how they trade, not by whether they can pay. No general launch date has been given.
What is vendor-reported, and what is missing
De¹ has not published an accuracy rate, a calibration test, a latency figure, a model size, a training-data description or a list of supported chains and venues. The only figures in circulation are about the platform, not the model, and all are De¹'s own:
- "Hundreds of millions" in daily trading volume used to label signals (press release, no exact number).
- More than 1,000 liquidity sources across more than 40 chains, as reported by DevCuration.
- A $500,000 investment from YZi Labs through its EASY Residency program, announced 28 August 2026, which DevCuration describes as a strategic investment with no disclosed valuation.
The funding shows backing, not performance. The same DevCuration piece points out that a blockchain can prove a trade happened but cannot show the trade was sound, and that it is unproven whether the model stays calibrated when markets are volatile. Most of the launch coverage reprints De¹'s own notice, so we found no independent verification.
How it compares on paper
| Model | What it decides | Access | Published accuracy |
|---|---|---|---|
| De¹ Ginza 1.0 | Likely execution outcome of an on-chain trade | Waitlist | None |
| Kakao Kanana-2-v-decide | Probability for each option you define, with image input | Open weights | See our explainer |
| Salesforce Koa Decisions | Pick, yes/no or score for routing | Salesforce products | Vendor-reported |
| A general chat LLM given a prompt | Anything, as generated text | Public APIs | Depends on your task |
The first two are general decision models you point at your own options. Ginza is narrower: one domain, one question, with training data most buyers cannot inspect. That specialization could be an advantage if its forecasts hold up. It also means the usual way to evaluate a decision model, asking it the same classification in both directions and checking known-answer controls, does not transfer cleanly, because a forecast of a future trade has no instantly checkable answer. A fair test would need a log of forecasts compared with realized fills over weeks. For the general method, see how we weigh structured-output reliability in chat models, which is what you would otherwise use for a pre-signing check.
Who should care
If you trade or run agents on-chain, apply to the waitlist and keep a written log of every forecast and the realized outcome before you let it influence real money. If you are choosing a decision model for routing, moderation or scoring, Ginza is not for you. Pick from the general models above and test on your own data.
We'll retest Ginza when it has a public endpoint or when De¹ publishes numbers we can check.