d1
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.4/10
Self-serve hosted API (console.liquid.ai, free d1:free tier) with public docs and a migration guide; works with TypeSafe's official SDKs through base_url. Live on OpenRouter since 01/10 (launched at $0.04/M, now $0.02/M input) and available through Vercel AI Gateway since 09/10. No SLA.
- Capability6.5/10
Full Noul / Choice / Score in one call, zero generated tokens. Calibration is claimed but no method or ECE is published, and there are no latency figures. The only benchmark is Liquid's internal run of HF's Decision Index 0.2.1; d1 is not on the official index.
- Adoption27/100
Big launch for the category (~1.8k likes / ~680k views on X within a day, press pickup), backed by an established lab. Now listed on OpenRouter (launch post ~160 likes / ~22k views). No production cases reported yet.
Vendor claims
- First model to outperform Jev on Hugging Face's Decision Index: 58.9 vs 57.9 (Liquid's internal reproduction of Decision Index 0.2.1)[Vendor claim — not independently verified]
- Area scores vs Jev 1.13: Arts 45.5 vs 37.7, Language 67.6 vs 62.0, Retrieval 60.7 vs 55.4, Tools 74.1 vs 75.1, Knowledge 43.3 vs 51.3[Vendor claim — not independently verified]
- Wins on multilingual evals, more robust against prompt injection, handles longer inputs more effectively[Vendor claim — not independently verified]
- Returns calibrated probabilities with zero generated tokens (usage.output_tokens is always 0)[Vendor claim — not independently verified]
News covering this model
d1 is the first decision model from Liquid AI, the company behind the LFM model family. Liquid announced it on 29 September 2026 (15:35 BRT) as “the first model to outperform Jev on @huggingface’s Decision Index”, with a hosted API at console.liquid.ai and OpenRouter listed as “soon”.
It follows the System One contract that Jev introduced: you send state plus named typed questions, and get back typed answers with probabilities. Liquid’s docs say d1 answers “in a single call with zero generated tokens”, and every example response shows usage.output_tokens: 0.
Specs
| Attribute | Value |
|---|---|
| Company | Liquid AI |
| Launch | 29 Sep 2026 (X announcement, 15:35 BRT) |
| Status | Public self-serve API; model id d1:free in the docs. Liquid does not call it GA |
| Endpoint | POST https://api.liquid.ai/decisions/v1/systemone |
| Auth | Bearer key from console.liquid.ai (keys start with liquid_) |
| Question types | noul (yes/no probability), choice (distribution over named options), score (probability-weighted position on an ordered rubric) |
| SDKs | TypeSafe’s own clients: pip install typesafe-sdk / npm install @typesafe-ai/sdk with base_url="https://api.liquid.ai" |
| Pricing | Liquid API: free d1:free tier. OpenRouter (since 01/10): $0.02 / M input, $0 output, 65K context (launched at $0.04 / M; Liquid’s pricing page and OpenRouter both list $0.02 as of 16 Oct 2026) |
| Weights | The hosted d1 checkpoint is still unpublished. On 7 Oct 2026 Liquid shipped a separate open family, d1-3B and d1-omni-600M. Those cards do not claim to be the hosted model |
| Parameters / base model | Not disclosed |
| Latency | No figures published (the 0–0ms frontmatter value means “no data”) |
On OpenRouter (1 October 2026)
@OpenRouter announced on 1 October 2026 (15:30 BRT) that d1 is live as liquid/d1: yes/no, choice or score questions with a probability for every option, “zero data retention” (OpenRouter’s wording; see the data note below), $0.04 / M input, $0 output, 65K context at launch (the price has since been cut to $0.02 / M). That delivers the “available on OpenRouter soon” from launch day and gives d1 its first third-party distribution. See the news.
Data handling: correction (17 Oct 2026)
“Zero data retention” comes from OpenRouter’s 1 Oct post, not from Liquid. OpenRouter’s current liquid/d1 page has no data-policy statement for d1; the line “may be retained and used to train Liquid models” that appears on that page belongs to other Liquid models in its sidebar (LFM2.5-Embedding-350M and LFM2.5-2.6B). Our 16 Oct radar note read it as being about d1; that was wrong. Liquid’s own sources: the Terms allow it to “retain Your Content and Usage Data as reasonably necessary to provide, secure, and improve the Services”, subject to service-specific or platform terms; the privacy policy lists improving AI models with “inputs and outputs you share”; the pricing page says it never trains on customer data in its enterprise tier; the d1 docs are silent. We cannot confirm zero retention on either route. Details in the news.
On Vercel AI Gateway (9 October 2026)
Vercel made Liquid d1 available through AI Gateway, including vision input, the Vercel AI SDK experimental_decide interface, and OpenAI-compatible Decisions API and TypeSafe-compatible routes. This adds a distribution and integration path; it does not change the hosted checkpoint or provide an independent benchmark. Catalog scores stay unchanged. Announcement.
TypeSafe-compatible by design
The request and response shapes match TypeSafe’s /v1/systemone, and Liquid’s quickstart uses TypeSafe’s official Python and TypeScript SDKs pointed at https://api.liquid.ai. Code written for Jev should move over by changing the base URL, key and model id. Liquid’s migration guide goes the other way too, showing how to replace LLM classification, routing, moderation and scoring calls with d1. A road-decider demo lives in Liquid’s cookbook repo.
The Decision Index claim
The launch chart is labelled “Internal reproduction of Decision Index 0.2.1 by Hugging Face”. It compares d1 with Jev 1.13:
| Area | d1 | Jev 1.13 | Diff | Benchmarks |
|---|---|---|---|---|
| Arts | 45.5 | 37.7 | +7.8 | 7 |
| Language | 67.6 | 62.0 | +5.6 | 10 |
| Retrieval | 60.7 | 55.4 | +5.3 | 6 |
| Tools | 74.1 | 75.1 | −1.0 | 5 |
| Knowledge | 43.3 | 51.3 | −8.0 | 10 |
| Index | 58.9 | 57.9 | +1.0 | 5 areas |
UnverifiedClaim. Liquid ran these numbers itself. The official Decision Index Space (v0.2.1 data, generated 28 Sep 2026) does not list d1. What we could check: the Jev column matches the official Jev 1.13.0 row (balanced skill 57.91; Knowledge 51.4, Language 62.0, Retrieval 55.4, Tools 75.1, Arts 37.7). So Liquid used the published Jev baseline and ran d1 on its own. The Index row is the index’s balanced skill score, not a plain average of the five areas.
Read together, the chart says d1 is ahead on arts and language tasks, level on tools, and clearly behind on knowledge and reasoning. The overall lead is one point. The claims about multilingual robustness, prompt injection and long inputs come without their own numbers.
Fit / anti-fit
Fit when you need: a hosted System One endpoint that is not TypeSafe, an easy A/B against Jev (same SDK), or multilingual text classification where Liquid claims an edge.
Anti-fit when you need: local or on-prem inference (no weights yet; Liquid replied “for sure” when asked about local runs), knowledge-heavy judgments (its weakest area on its own chart), published latency or pricing guarantees, or independently verified accuracy.
Limits
- No public SLA. Liquid’s own API still only shows
d1:free; the first price was OpenRouter’s ($0.04 / M input at launch; $0.02 / M by 16 Oct, also on Liquid’s pricing page). - Parameters, architecture, base model and training method of the hosted d1 are not disclosed. The open d1-3B and d1-omni-600M (7 Oct 2026) are a different release, with published bases (LFM2.5-VL-3B and LFM2.5-Encoder-350M). They are not a drop-in for
api.liquid.ai. - “Calibrated probabilities” is a vendor claim with no published ECE or method.
- Every benchmark number above is Liquid’s own run (UnverifiedClaim).
Why it is in the catalog
Liquid presents d1 as its own trained decision model, not as a general LLM behind a decision wrapper. It has a public, documented typed API that returns probabilities. That passes the catalog criteria the same way hosted Jev does. Compare the OpenAI Decisions API, which OpenAI describes as a way of using GPT-6 Luna and which has no public docs; it stays under runtimes.
Artificial Analysis Decision Index v1.0 (13 Oct 2026)
Per the AA launch article, d1 has the top AA-Omniscience-Decision accuracy (41) among decision models but a non-hallucination rate of 2 (it almost never picks “I don’t know”). Its index score is shown only in the chart, not in the text, so we don’t quote it and scores are unchanged. AA is a separate, independent board (four evaluations: AA-Omniscience-Decision, HLE-MultiChoice, When2Call, WildGuardTest). Its scores and ranks are not comparable with Decision Index 0.3 or JevBench. AA finds every decision model overconfident overall, mainly on the knowledge and reasoning tasks; most are fairly well calibrated on When2Call and WildGuardTest.
Links
- Announcement on X
- Liquid docs: Decision Models
- Liquid docs: Decision Model Guide (migration)
- Vercel AI Gateway changelog
- Cookbook: road-decider demo
- Hugging Face Decision Index (official Space)
- News: Liquid AI launches d1 · d1 goes live on OpenRouter
- Compare: Jev
- Bigger sibling: d1-ultra (15 Oct 2026; #1 on the AA Decision Index v1.0)
