Data as of October 27, 2026 · How the score is built
d1-2B
Decision readingd1-2B is tracked, but not publicly ranked yet. The profile exposes 0 sourced benchmark rows and leaves unsupported fields blank until a published record exists.
Released Oct 27, 2026 — see all recent releases
Lineage
The sequence follows explicit supersedes links. Each score is estimated for that model; a relative can inform a sparse estimate but never sets a floor, so a newer release can score below an earlier one. Scores and prices remain blank when the corresponding public row or first-party rate is unavailable.
Spec sheet
Each documented value carries its source. Missing fields stay visible as not sourced or not published, rather than disappearing from the page.
- API model ID
- LiquidAI/d1-2BLiquid AI d1-2B model card
- Context window
- 128KLiquid AI d1-2B model card
- Maximum output
- Not sourced yet
- Knowledge cutoff
- Not sourced yet
- Input modalities
- Not sourced yet
- Output modalities
- Not sourced yet
- Parameters
- Not sourced yet
- Availability
- It has 1.92B parameters and reads text, pictures and video within 128K tokens; Liquid says answers get less reliable past about 64K tokens of state. Liquid reports a 5.7 ms single-question call on an NVIDIA H100 PCIe and 3.8 ms on an AMD MI325X (warm bf16 calls after model.compile(), median of 20); these are latencies, not token throughput. It is not a chat model. The weights license has commercial-use conditions.
- Cloud regions
- Not tracked yet
- Lifecycle
- Current
- API capabilities
- Tool calling, structured outputs, and batch support are not tracked yet
- Prompt caching
- Not documented in the pricing recordLiquid AI d1-2B model card
- Self-host
- Open weights available; hardware estimate not sourced
- Rate limits
- Not tracked yet
How to read this profile
The visual layer above carries the decisions. These notes preserve the model, ranking, coverage, and family context behind the numbers.
We track d1-2B, but no weighted text-model benchmark result is published on the site yet. This page shows the metadata and separate protocol evidence we can verify now; a BenchLM score will appear only if compatible public evaluations land.
d1-2B is a open weight model with a 128K context window. No explicit reasoning mode is documented in this profile.
It has 1.92B parameters and reads text, pictures and video within 128K tokens; Liquid says answers get less reliable past about 64K tokens of state. Liquid reports a 5.7 ms single-question call on an NVIDIA H100 PCIe and 3.8 ms on an AMD MI325X (warm bf16 calls after model.compile(), median of 20); these are latencies, not token throughput. It is not a chat model. The weights license has commercial-use conditions.
Liquid AI released d1-2B on October 27, 2026, with the LFM3 family. It is built on LFM3-VL-2B, has 1.92B parameters and a 128K context, reads a text or JSON state with pictures or video, and answers any number of named questions in one forward pass with probabilities and no generated tokens. Beyond yes/no, choice and score answers it adds Point (boxes and counts), Read (extracted spans, or not found), Estimate (a number with a range) and Pick (multi-select). Liquid reports 51.1 on the Decision Index public suite (v0.3), 85.2 on Decision Bench v4 (decisionbench.ai), 77.1 on DecisionBench v1 (Hanno Labs), 77.9 on JevBench v1.4.1, 75.5 on ImajevBench v2.0-lite, 61.3 on Typed Decisions and 63.2 on its own board of over a hundred public and private benchmarks, against 49.0, 83.6, 73.1, 76.2, 64.0, 65.0 and 58.8 for d1-3B. Liquid scores all four open d1 checkpoints on the same items with the same scoring, so some d1-3B and d1-omni-600M figures differ slightly from their own cards. These provider-run decision results are not stored on Decision Index 0.2.1, Decision Index 0.3, JevBench or Decision Index v1.0 lanes, and they do not establish general LLM capability rankings.
d1-2B sits in the Liquid d1 open family with d1-Omni-2B, d1-3B, d1-omni-600M. Its explicit predecessor is d1-3B. The profile has no source-displayable benchmark row yet.
Last updated October 27, 2026. Runtime fields remain blank until a sourced snapshot exists.
Deployment options
Self-host and provider-specific paths stay separate from benchmark evidence so operating constraints are visible before a score becomes the whole decision.
Published weights are available, but BenchLM does not yet have a sourced parameter and VRAM profile for this exact model. Hardware cost estimates stay unavailable until that sizing record is complete.
Estimate VRAM from known parametersQuestions
How does d1-2B perform overall in AI benchmarks?
d1-2B does not have any source-displayable benchmark rows yet, so this profile does not assign a public score or rank. Documented specifications remain visible, while score-led charts and claims stay unavailable until a published evaluation can be attached to the exact model.
Is d1-2B open source?
d1-2B is an open-weight model from LiquidAI. Its weights can be downloaded for local or hosted deployment, subject to the published license. Open weight does not automatically mean open source: training data and training code may remain private, and commercial restrictions can still apply.
Which sibling models are related to d1-2B?
d1-2B belongs to the Liquid d1 open family. Related tracked variants include d1-Omni-2B, d1-3B, d1-omni-600M. A sibling link indicates shared lineage or a documented configuration relationship; it does not mean the variants have identical pricing, context limits, benchmark evidence, or deployment behavior. Compare before switching.
Does d1-2B have full benchmark coverage on BenchLM?
No. d1-2B currently has 0 source-displayable rows. The profile exposes published, non-generated evidence and leaves missing categories blank until an exact evaluation is available. Coverage describes how much was measured; it is not a penalty added to an individual benchmark result.
What is the context window size of d1-2B?
d1-2B has a documented context window of 128K. That figure is the maximum combined prompt and retained-conversation space reported for this exact model; it is not the maximum output length. The profile keeps output limits separate because providers often publish those limits independently.