What is d1-omni-600M?
d1-omni-600M is a small open-weight decision model from Liquid AI, released on Hugging Face on 7 October 2026 alongside the larger d1-3B. It writes no text. You pass a state (a string, JSON, images or one voice clip of up to 30 seconds) and a set of named questions, and it returns a yes/no probability, a pick from named options or a score on an ordered scale, read straight from the model's output distribution with zero output tokens.
It has 587 million parameters: a 381M shared trunk and decision head built on the bidirectional LFM2.5-Encoder-350M, a 94M SigLIP2 vision encoder and a 112M FastConformer audio encoder. The context is 16,384 tokens, but with images the text part is cut to 896 tokens, and a request can carry images or audio, not both. Liquid calls it an early research checkpoint and publishes no speed figures for it.
The weights are under the LFM Open License v1.0, which withholds commercial rights from organisations with annual revenue of $10 million or more. On 27 October Liquid released d1-Omni-2B as part of LFM3, and its own card says the newer model scores above d1-omni-600M on every benchmark.
Is d1-omni-600M worth it?
d1-omni-600M is notable for what it accepts, not for how well it answers. Few decision models take a voice clip, and fewer do it in under 600 million parameters. Accuracy is the problem. Liquid's own card puts it at 15.95 on Decision Index 0.2.1, against 48.57 for d1-3B and 28.97 for the 2.3B Decider 2B. The community Jev Decision Index 0.4.1, run on one RTX PRO 6000 and refreshed on 29 October, scores it 11.75 and places it 228th of 247 entries, behind vLLM Semantic Router's d3-edge (17.69) at the same size.
The text-only numbers on the card are kinder because they are narrower: 95.8 on Civil Comments toxicity and 79.5 on PAWS-X paraphrase, both the best in Liquid's comparison table, and a mean of 78.4 over seven public tasks against 77.1 for Decider 2B. Those are single-skill classification jobs, the kind a fine-tuned encoder has always done well. Broad decision work, with knowledge and multi-step rules, is where it falls apart: 8.3 on the index's knowledge area and 6.8 on arts.
Calibration is weak too. On the community index its expected calibration error is 0.128, against 0.047 for d1-3B, so its probabilities should not be used as thresholds without your own checks. Liquid says image and audio answers are the raw softmax, without the per-type temperatures applied to text.
The more important fact is its successor. d1-Omni-2B, released on 27 October, takes the same three modalities and Liquid reports it at 49.0 against 18.3 on the Decision Index v0.3 public suite and 75.3 against 58.0 on JevBench v1.4.1. The 600M model remains useful as a reference for how small an audio-capable decision model can go, and for experiments on hardware where 2 billion parameters will not fit.
Where it earns its place
An embedded engineer building a voice-controlled appliance prototype
Route short spoken commands to one of a few intents offline
One 16 kHz clip and a choice question return an intent with probabilities in a single pass. Audio training covered English speaker-to-assistant requests only, so test your own commands, and move to d1-Omni-2B if the 587M model's error rate is too high.
A trust and safety engineer at a small forum
Flag toxic comments before they are posted, on a CPU server
Toxicity is the one job where Liquid's table puts it first (95.8 on Civil Comments). A yes/no question per comment needs no generation, and float16 kept the same top answer as float32 on all 243 text rows Liquid checked.
Before you commit
- Setup
- pip install transformers 5.15 or later, torch, torchvision, pillow and soundfile, then load with trust_remote_code=True and call model.system_one(state, questions). Avoid bfloat16, which changed the top answer on 0.8% of text and 1.7% of audio rows in Liquid's checks.
- Integrations
- Questions follow the Decision Index schema (noul, choice and score), the same shape Jev and pplx-decider clients send; system_one_batch handles many requests per call. Liquid's hosted d1 API serves the newer models, not this one.
- Data and privacy
- Runs locally, so nothing leaves your machine. The LFM Open License v1.0 grants commercial use only to organisations under $10 million in annual revenue and to qualifying non-profits.
- Cost at scale
- Free to run; at 587M parameters it fits on a laptop CPU or small GPU. Liquid publishes no hardware or speed figures for this checkpoint; the community index measured a 10.9 ms median on an RTX PRO 6000.
- Limits
- 16,384-token context, cut to 896 tokens of text when images are attached; images or audio per request, not both; audio clips stop at 30 seconds. Expected calibration error of 0.128 on the community index.
What’s good
- Text, images or a 30-second voice clip in one 587M model
- 95.8 on Civil Comments toxicity (self-reported)
- 10.9 ms median per request on the community index
- Zero output tokens; answers come with probabilities
What’s not
- 11.75 on the community Decision Index 0.4.1, 228th of 247
- Superseded by d1-Omni-2B on every benchmark Liquid reports
- Image and audio answers are not calibrated
- No commercial licence at $10M+ annual revenue
Best for
Researchers and hobbyists testing how far an audio-capable decision model can shrink, not teams choosing a production classifier.
d1-omni-600M alternatives
d1-omni-600M: common questions
- Is d1-omni-600M worth it?
- A curiosity rather than a component: tiny and fast, but it ranks 228th of 247 on the community decision index and its own maker has already replaced it. We rate it 3.0 out of 5.
- How much does d1-omni-600M cost?
- Free weights under the LFM Open License v1.0; commercial use needs a licence from Liquid AI at $10 million or more in annual revenue. No hosted endpoint for this checkpoint.
- What are the downsides of d1-omni-600M?
- 11.75 on the community Decision Index 0.4.1, 228th of 247; superseded by d1-Omni-2B on every benchmark Liquid reports; image and audio answers are not calibrated; no commercial licence at $10M+ annual revenue.
- Who is d1-omni-600M best for?
- Researchers and hobbyists testing how far an audio-capable decision model can shrink, not teams choosing a production classifier.
- How much does d1-omni-600M cost for a team or in production?
- Free to run; at 587M parameters it fits on a laptop CPU or small GPU. Liquid publishes no hardware or speed figures for this checkpoint; the community index measured a 10.9 ms median on an RTX PRO 6000.
- Is d1-omni-600M safe for company data?
- Runs locally, so nothing leaves your machine. The LFM Open License v1.0 grants commercial use only to organisations under $10 million in annual revenue and to qualifying non-profits.
- What are the best d1-omni-600M alternatives?
- The closest alternatives we have reviewed are LFM3, Perplexity Decider (pplx-decider), d3-turbo (vLLM Semantic Router) and Jev.
Answers checked Oct 31, 2026.