Liquid AI joins the decision-model wave
D1 API ($0.04/M input) and open-weight d1-3B and d1-omni-600M for edge devices
Confirmed
The takeaway
On Oct 5, 2026 Liquid AI launched d1, a hosted “decision model” that answers typed questions (yes/no, label choice, scale scores) with probabilities instead of generated text, billed at $0.04 per 1M input tokens with no output charge.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 10 October 2026
Your AI and this story
- GPT-6 Astra158 days after its cutoff
- Claude Opus 5.597 days after its cutoff
- Gemini 3.8 Flash188 days after its cutoff
- Grok 4.7127 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 97 days before it.
Key facts
- d1 (Oct 5, hosted): text and images; three question types (boolean, single-label classification, probability-weighted scale scoring); $0.04 per 1M input tokens, no output-token charge; images count 1.5 tokens per 32x32 patch (1,536 tokens for 1024x1024) (Liquid AI)
- Liquid AI claims d1 ‘matches or beats GPT-6.1 Sol’ on four of six real-world applications it tested, at 19x to 200x lower cost than GPT-6.1 Sol and Claude Opus 5.5; 200-300 ms text decisions; 85-97% accuracy on production-line visual defect inspection (company figures)
- d1 hosted access: Liquid AI Console, Vercel AI Gateway, OpenRouter (text only at launch, vision ‘coming soon’)
- Oct 7 open release: d1-3B (built on LFM2.5-VL-3B with a 400M SigLIP2 NaFlex vision encoder, 32,768-token context) and d1-omni-600M (text+vision or text+audio); Hugging Face repos LiquidAI/d1-3B, LiquidAI/d1-omni-600M plus GGUF and w8a8 builds; license ‘lfm1.0’ (LFM Open License v1.0)
- d1-3B: 48.57 on the Decision Index v0.2.1 public split, ahead of every model under 10B and above Perplexity’s Decider 35B-A3B (47.11); mean 82.9 on seven public text datasets and 74.1 on eleven image datasets (Liquid AI, self-reported)
- d1-3B latency per question: 8 ms on an RTX 4090, 9 ms on AMD MI325X, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin, 50 ms on Jetson Orin Nano; 384px images in under 18 ms on GPUs; native llama.cpp support on Apple, AMD, Qualcomm and NVIDIA
- d1-omni-600M scores 15.95 on the same Decision Index split (secondary report)
What happened
Liquid AI, maker of the LFM (“Liquid Foundation Models”) family of small models, announced d1 on Oct 5, 2026. Like TypeSafe’s Jev, d1 does not write text: given a question and a fixed answer format it returns a calibrated answer in one forward pass, and customers pay only for input tokens ($0.04 per million). It handles yes/no questions, choice among named labels and rubric-style scores, and accepts images. Liquid AI compared it with GPT-6.1 Sol and Claude Opus 5.5 on six of its own application tests and reported equal or better results on four, at a small fraction of the cost. These are company figures.
Two days later it released open weights for two small versions meant to run on devices: d1-3B (built on LFM2.5-VL-3B) and d1-omni-600M, which also takes audio. Liquid AI reports 8 ms per decision for d1-3B on a consumer RTX 4090 and 50 ms on a Jetson Orin Nano.
Why it matters
In two weeks most of the AI industry shipped “decision models”: TypeSafe’s Jev, then Cloudflare’s Clef, Perplexity’s Decider, Liquid AI’s d1, OpenAI’s Decisions API and Microsoft’s Decision-1. Liquid AI’s versions are the smallest. They bring the format to phones, robots and factory cameras, where a full LLM call would be too slow or expensive.
Sources
4 sources from 3 sites. Numbers match the chips in the text.
4 sources: 3 primary, 1 press
Primary
- Liquid AI: Introducing d1, the most capable decision model, now with vision (Oct 5)liquid.ai, official
- Liquid AI: Open d1, edge decision models for text, vision and audio (Oct 7)liquid.ai, official
- Hugging Face: LiquidAI/d1-3Bhuggingface.co, code
Press
- AI Weekly: Liquid AI opens d1-3B and d1-omni-600M decision modelsaiweekly.co, press
Changes
- Filed (found via the nocode.mba October release tracker; Liquid AI blog and Hugging Face read directly)