d1 on Ollama: Liquid AI's 3B decision model, hands on

Liquid AI has released two decision models on Ollama today: d1, a 3B model that claims the best under-10B decision accuracy, and d1-omni, a 600M edge model that also hears. Here's what each one needs.

Listen to this post · 12 min

Accessibility: the whole post is read aloud — Space plays or pauses, ← → skip 10 seconds, and you can change the speed. It keeps playing with your screen locked.

Share this post

In short

  • d1 is Liquid AI's 3B decision model on Ollama from 27 October 2026 — it returns calibrated probabilities instead of generating text, and its 3.7GB download is small enough for a laptop.

Liquid AI has put two decision models on the Ollama library today: d1, a 3B model built on LFM2.5-VL-3B, and d1-omni, a 600M multimodal sibling built on LFM2.5-Encoder-350M. Like the other decision models that have landed this month, they don't chat and don't write — you give them a state and typed questions, and they return calibrated probabilities in a single forward pass. d1's pitch is the best under-10B accuracy at a size that fits a laptop. d1-omni's pitch is that it's tiny, and it hears.

What is d1? A decision model, not a chat model

d1 is a 3B decision model from Liquid AI, built on LFM2.5-VL-3B, and it's on the Ollama library as of today, 27 October 2026.

It doesn't chat and doesn't generate text. You give it a state — text, JSON, images, or a mix — plus up to 64 typed questions, and it returns a calibrated probability for every answer in a single forward pass. Zero output tokens. The response to a support ticket might be: urgent, 0.981 true; team, technical at 0.973; severity, 2.58 on a 0–3 scale. No prose to parse, no prompt engineering to get the model to answer in the format you want.

It's the latest in this autumn's rush of decision models on Ollama — we've covered kev, pplx-decider, Clef Flash and Strands Decider over the past fortnight, and d1 is the first from Liquid AI. The readme says it's compatible with the Jev and System One APIs, so it works through Ollama's /v1/systemone endpoint and ollama run --ask.

The headline number: Liquid AI reports 48.57 on Decision Index 0.2.1 — the best score for a decision model under 10B, ahead of the 35B-A3B Decider at 47.11. That's vendor-reported, as always with release-day numbers; we haven't seen independent benchmarks yet.

It's also multilingual: Arabic, Chinese, English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Thai and Vietnamese.

d1: sizes, memory and hardware

Two tags, and the naming is worth a careful look because it's counter-intuitive:

TagDownloadContextInputs
d1:latest (same as d1:3b)3.7GB32KText, Image
d1:2b3.9GB125KText, Image

Note that the 2b tag is the larger download (3.9GB vs 3.7GB) and the one with the long context window — 125K against the 3b's 32K. If you're judging long documents, the smaller-parameter model is the one you want.

Our usual rule of thumb: budget roughly the download size in RAM or VRAM, plus headroom for context. At 3.7–3.9GB of weights, that puts d1 comfortably on an 8GB gaming GPU, a 16GB Mac, or a modern laptop with 16GB RAM. Both tags accept text and image inputs.

For context on where that sits in this month's releases: pplx-decider's 27B variant was a 17GB download when we covered it, and Clef Flash's 9B was 11–12GB. At 3.7GB, d1 is the lightest of the decision models we've looked at this month, and it isn't restricted to the Apple Silicon MLX runner the way kev and Strands Decider were on release day.

Running d1: CLI and API

Pull it, then ask:

ollama pull d1
ollama run d1 --ask "Is a refund being requested?" < ticket.txt

The state can also follow the model name instead of coming from stdin.

Choice questions — you get the chosen option plus a probability for each:

ollama run d1 --ask "Which team should handle this?" \
  --choice billing="Payments and refunds" \
  --choice technical="Outages and errors" < ticket.txt

Score questions — a probability-weighted level, numbered from 0:

ollama run d1 --ask "How urgent is this?" \
  --level "Routine" \
  --level "Soon" \
  --level "Immediate" < ticket.txt

The API is Ollama's /v1/systemone endpoint. Put the text you want judged in state, your questions in questions (1 to 64, named, answered in the same order), and optional base64 PNG/JPEG/WebP images shared by all questions:

curl http://localhost:11434/v1/systemone -d '{
  "model": "d1",
  "state": "Our app crashes on launch for every user since this morning'\''s update. Please fix it now.",
  "questions": {
    "urgent": {"type": "noul", "instructions": "Is this support request urgent?"},
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "billing": "Payments, invoices, and refunds",
        "technical": "Outages, errors, and configuration"
      }
    }
  }
}'

Three question types: noul (true/false with a probability), choice (an object mapping options to descriptions), and score (2–10 level descriptions, lowest first).

One thing worth internalising: the confidence field runs 0–1 and shows how strongly the probabilities favour one answer. It is not the chance the answer is right. A confidently wrong answer is still possible.

For Python, TypeSafe's official SDK works against Ollama. The SDK requires an API key, but Ollama ignores it, so any value works — set TYPESAFE_BASE_URL=http://localhost:11434 and TYPESAFE_API_KEY=ollama.

As with the other decision models this month, the Ollama library page is the only documentation we've seen — no LM Studio, llama.cpp or vLLM instructions as of release day.

What d1 is good for — and what it can't do

The readme's use-case table maps well onto things people currently do badly with chat models:

  • Routing and triage — define the destinations and when each applies; get the chosen destination and each one's probability.
  • Moderation — express the policy as a yes/no question; get true or false plus the probability of each.
  • Judging answers — a rubric of levels; get the probability-weighted level and each level's probability.
  • Inspecting images — typed questions about what's in the frame, answered with probabilities.
  • Gating agent steps — a condition and the agent's proposed action; go or no-go, with probabilities.

Anywhere you're currently parsing a chat model's prose to make a binary call, this is the cleaner tool.

The limits are the flip side of the design:

  • It never generates text. No summaries, no drafts, no chat. If you want prose, you still need a chat model alongside it — something small like qwen3:1.7b or llama3.2 pairs nicely if you'd rather call than host.
  • The 3b tag's 32K context limits how much state you can judge in one request; the 2b tag's 125K window is the pick for long documents.
  • All quality numbers are Liquid AI's own. No independent benchmarks yet — par for the course on release day, but worth waiting on before you trust it for anything critical.

d1-omni: a 600M decision model that also hears

The second half of the family is more unusual. d1-omni is a multimodal 600M decision model built on LFM2.5-Encoder-350M — 587M parameters total, with small vision (94M) and audio (112M) encoders on the same trunk.

TagDownloadContextInputs
d1-omni:latest (same as 600m)850MB125KText, Image
d1-omni:2b4.4GB125KText, Image

At 850MB, the 600m is the smallest decision model we've covered on Ollama, and the first with audio input. An 850MB download is well within reach of modest hardware — old laptops and small boards included — though the page doesn't name specific devices, so treat that as our read of the size rather than a vendor claim.

Speech works like this: up to 30 seconds of English speech per request, sent as a base64 WAV in the images field. Send audio or images, not both — clips are cut at 30 seconds. The audio training data was English requests from a speaker to an assistant, so it works best on questions about what kind of utterance it is, its topic, and what the speaker wants. Don't expect it to answer arbitrary questions about the audio.

Images: keep text short when sending them — the model was trained with up to 896 tokens of text alongside images, so trim the state and questions.

Liquid AI's own guidance is clear: use d1 for the most accurate decisions over text and images; use d1-omni when size matters or you need audio.

It runs the same way:

ollama run d1-omni --ask "Is this support request urgent?" < ticket.txt

…or via /v1/systemone. Licence: LFM Open License v1.0.

d1 vs d1-omni: which one do you want?

d1d1-omni
BaseLFM2.5-VL-3BLFM2.5-Encoder-350M
Downloads3.7GB (3b) / 3.9GB (2b)850MB (600m) / 4.4GB (2b)
Context32K / 125K125K
InputsText, ImageText, Image, Audio (30s English)
Best forAccuracy over text and imagesWhen size matters, or audio

Liquid AI's benchmark table (public benchmarks framed as decisions, vendor-reported) splits the wins:

Benchmarkd1-omnid1
Civil Comments95.893.0
MASSIVE intent86.187.3
PAWS-X79.576.9
BoolQ77.786.7
SQuAD 2.074.085.3
Mean (7 benchmarks)78.482.9

d1-omni actually beats d1 on Civil Comments and PAWS-X. But on the harder Decision Index 0.2.1, Liquid AI reports 15.95 for d1-omni vs 48.57 for d1 — a big gap. Don't use the 600m where accuracy matters.

Rule of thumb: d1-omni for small devices, audio, or cheap high-volume triage; d1 for anything where the decision needs to be right.

How d1 fits the decision-model landscape

Decision models have arrived in a rush this October. In the fortnight before today's release we covered Strands Decider (a 2B model, Mac-only MLX on release day), Clef Flash (Cloudflare's 9B, 11–12GB), kev (a family of checkpoints from 812MB to 18GB, also MLX) and pplx-decider (Perplexity's 27B, 17GB) — plus nimble-v3, a 9B from Bespoke Labs.

Against that field, d1's pitch is the best under-10B accuracy (48.57 on Decision Index 0.2.1, per Liquid AI) at a 3.7GB download — a combination none of the others offered. The bigger models needed bigger hardware; the MLX ones needed a Mac. And d1-omni brings something none of them had at all: audio input, at 850MB.

One practical note: the same /v1/systemone endpoint and --ask CLI work across these models, so pipelines built for one transfer fairly easily — but check whether the others' question formats match before assuming.

Getting started

One command, no prompt engineering:

ollama run d1 --ask "Is this support request urgent?" < ticket.txt

If you have a spare GPU or an idle Mac, decision models are cheap to host — no output tokens means fast responses — and the Inpherio connector works with Ollama. New hosts pay no platform fee for 12 months, up to £250 of earnings; details at providers.

And before trusting the vendor numbers for anything critical, watch for independent benchmarks. We'll be looking out for them.

Questions people ask

How do I run d1 locally?

Install Ollama, then ollama run d1 --ask "your question" < state.txt. For structured questions use --choice or --level flags, or POST to Ollama's /v1/systemone endpoint.

Models in this post — live on Inpherio

Have a GPU, a Mac or a spare PC?

Let it earn while you sleep.

Run open models for people who need them. You set the price and keep everything but the 15% fee.

Become a host

Building with AI?

Every open model, one API.

OpenAI-compatible. Priced by the people who run the hardware.

Start free See the marketplace

Read “d1 on Ollama: Liquid AI's 3B decision model, hands on” at https://inpherio.co.uk/blog/d1-on-ollama-liquid-ai-s-3b-decision-model-hands-on