LFM3-Omni-2B vs Moondream 3: One Benchmark in Common, Two Different Jobs

Put LFM3-Omni-2B and Moondream 3 side by side and the spec sheets look like siblings. Both are small multimodal models aimed at phones, laptops, cameras and robots. Both carry a 32K context. Both put roughly two billion parameters to work on every token. Both ship open weights you can download today. Search for a head-to-head between them and you get nothing, which is exactly why the pairing keeps getting made.

Then you read the two model cards carefully and the resemblance falls apart in a way that is easy to check. The only benchmark both cards report is DocVQA, where LFM3-Omni-2B scores 92.47 and Moondream 3.1 9B A2B scores 88.6. Everything else they measure is disjoint. One model returns spoken sentences. The other returns bounding boxes. They are not two answers to the same question.

A note on names, because the directory is confusing. Moondream 3 Preview shipped on 18 September 2025 and Moondream 3.1 9B A2B on 7 July 2026; every Moondream number in this piece comes from the 3.1 card, which is the checkpoint you can download. LFM3-Omni-2B was published on 27 October 2026 as one of six models in Liquid AI's LFM3 family. Both are brand new in the only sense that matters here: neither has been evaluated by anyone outside the company that built it.

What each model actually returns

The deepest difference is not parameter count. It is the output contract.

Moondream 3.1 exposes five skills with fixed shapes: query, caption, detect, point and segment. Detection returns at most 200 boxes, pointing at most 64 points, segmentation at most 50 COCO-RLE masks, all in normalised 0–1 coordinates as JSON. It is a spatial measurement instrument with a narrow, predictable interface. You can wire its output straight into a UI automator or a robot planner without an LLM parsing layer, because the output is already structured.

LFM3-Omni-2B has no such contract. It takes text, images, video with or without its soundtrack, and 16 kHz audio, and it generates text, 24 kHz speech, or both interleaved. There is no coordinate output path. If you want a location out of it, you ask in prose and parse the answer, which is a different reliability class from a schema that guarantees a box.

That single difference decides more deployments than any benchmark row. Swapping one model ID for the other is not a migration. It is a rewrite of whatever consumes the response.

The spec sheet, side by side

• Parameters — 2.17B total for LFM3-Omni-2B, all of it resident and computing vs 9.0B total for Moondream 3.1 with 2.0B active per token.

• Memory at rest — about 4.4 GB in BF16 by arithmetic on 2.17B parameters vs roughly 18 GB in FP16, about 6 GB in community INT4, and a 10 GB NVIDIA floor for the shipped fp8 build.

• Inputs — text, images, video at 2 fps with or without audio, and 16 kHz speech vs images and text only.

• Outputs — text, 24 kHz speech, or interleaved vs text, JSON, normalised coordinates and masks.

• Context — 32,768 tokens counting input plus output vs 32,768 text tokens plus up to 12 image tiles, around 2,304 vision tokens.

• Vision budget — up to 2,048 tokens per image at native resolution vs adaptive tiling to 12 crops at 384–1024 px.

• Licence — LFM Open License v1.0, commercial use licensed only below $10M annual revenue vs Moondream Model License 1.0, commercial at any size but no general-purpose hosted service.

• Published — 27 October 2026 vs 7 July 2026, with the 3 Preview branch dating to 18 September 2025.

The memory row is the one people skip, and on an edge device it is the whole argument. Moondream's 2B active parameters make it as cheap as LFM3-Omni-2B per token, but its 9B total makes it up to four times larger in memory at the same precision. Active parameters govern compute. Total parameters govern whether the thing fits. A model that runs at the same speed but needs 18 GB is not competing for the same hardware slot as one that fits in 4.4 GB.

The only benchmark they both report

DocVQA is document question answering, scored by ANLS, and on the two model cards it reads 92.47 for LFM3-Omni-2B against 88.6 for Moondream 3.1. The omni model is ahead by 3.87 points on the vision specialist's own turf.

That result should be read with three reservations, all of them load-bearing. Both figures are self-reported by the labs that built the models. Neither lab published its evaluation harness, prompt format or sample selection. And a 3.87-point gap on a benchmark with this much variance between harnesses is suggestive rather than decisive. What the row does establish is narrower and still useful: LFM3-Omni-2B does not fall apart on dense document text, which is the task you would expect a 2.17B any-to-any model to lose.

Do not extend the same reading to grounding. LFM3-Omni-2B reports RefCOCO at 73.55 and Moondream 3.1 reports RefCOCO-M at 87.1, and those are not the same measurement. Liquid's card notes that boxes are read in each model's native coordinate format; Moondream reports F1 at an IoU threshold of 0.5. Subtracting one from the other produces a number that means nothing. The honest statement is that Moondream's card is built around grounding and Liquid's treats it as one row among seventeen.

Where Moondream 3 is not close

Some of that is a reporting choice rather than a capability gap, and it is worth being precise about which. A 2.17B omni model can almost certainly locate a rough region if you ask it in words. What it cannot do is return 200 boxes in a schema with a bounded error rate, because nothing in its training or its output format was built for that. Moondream's entire product is the guarantee that the coordinates come back in a parseable shape. That guarantee is the thing you are buying, and no benchmark row from another lab measures it.

The same logic runs in reverse for latency. Moondream ships Photon, an inference engine that compiles the model into a single GPU program, and the vendor claims about 34 requests per second at batch 16 on an H100 with a 59 ms P50. Those are vendor figures, unreproduced, but the engineering behind them is specific to one model. LFM3-Omni-2B has no equivalent serving story yet.

Where LFM3-Omni-2B is not close

Moondream 3.1 has no audio path. Not a weaker one — none. Its input is an image and text, full stop. So every audio row on the Liquid card is uncontested: English ASR word error rate 6.36, multilingual ASR 6.72, English TTS UTMOSv2 3.80, VoiceBench 64.73, MMAR 51.40, and an internal tool-calling eval at 52.20.

The video-with-soundtrack numbers are the ones worth singling out, because they are the reason the model exists. Daily-Omni 70.09, WorldSense 47.10, AV-SpeakerBench 46.79 — all measured on video with its audio track, multiple-choice accuracy, all self-reported. Liquid's claim is that these beat Qwen2.5-Omni-7B and models up to 10.7B parameters, and on its own table they do: Qwen2.5-Omni-7B scores 61.32, 45.37 and 43.24 on the same three rows.

One detail on that card deserves more attention than it got. Liquid reports the same vision benchmarks twice — once with the question typed, once with the question spoken. RealWorldQA drops from 66.54 to 64.68, MMStar from 55.67 to 51.79, POPE from 89.34 to 83.78. Feeding the question through the audio encoder costs between two and six points depending on the task. That is the price of the speech path, stated plainly, and it is a more useful number than any single headline score.

What happens to Moondream 3 on 13 November

This is the part of the comparison that has a deadline. Prime Intellect announced it was acquiring Moondream on 13 October 2026, and Moondream confirmed it on 14 October. The post states that on 13 November 2026, thirty days after the announcement, the Moondream Cloud — both hosted inference and the Lens fine-tuning product — shuts down. Unspent purchased credits are being refunded, and customers are being directed to the acquirer's own inference platform.

Three things follow, and only the first is comfortable. The open weights survive and the licence is unchanged, so self-hosting is unaffected. Photon runs locally at no cost on version 1.3.0 and later with no API key. What does not survive cleanly is Lens: fine-tuned checkpoints and LoRAs live server-side, and there is no documented export path. If your product depends on a Lens fine-tune, the migration is not a configuration change, and sixteen days is not much time to discover that.

There is a smaller oddity worth flagging for anyone checking sources this week. Moondream's pricing page still lists Cloud plans, per-token rates and a $350/month Team tier, and says nothing about the shutdown. The announcement lives on a different page. A reader who checks only the pricing page on 28 October would not know the service ends in sixteen days.

What each one costs to run

Moondream publishes token rates: $0.30 per million input tokens and $1.00 per million output on the real-time tier, halved to $0.15 and $0.50 on batch. LFM3-Omni-2B publishes no hosted rate at all. Liquid says the family is available on its Playground and API, but the launch post carries no price, and the model card carries no price. That asymmetry is the cost story.

Run the Moondream arithmetic on a realistic call. A single image costs up to 2,304 vision tokens at full tiling; add a 50-token prompt and 100 output tokens and one call is 2,354 input tokens and 100 output tokens, or about $0.0008. At 10,000 calls a day that is roughly $8 a day, or $242 a month, dropping to about $121 a month on the batch tier. Whether that is cheap depends entirely on what you would otherwise spend on hardware, and that is the comparison LFM3-Omni-2B forces you to make because it gives you no per-token number to weigh against it.

On your own hardware the arithmetic reverses. LFM3-Omni-2B is about 4.4 GB of weights in BF16 before activations, the audio decoder and the KV cache — arithmetic from the parameter count, not a measured footprint, since the card publishes no memory figure. Moondream 3.1 wants a 10 GB NVIDIA card for its shipped build, or about 6 GB with a community INT4 quantisation. One model runs on a phone. The other runs on a Jetson Orin or better.

OrcaRouter is worth a sentence here only because the honest version is narrow: neither LFM3-Omni-2B nor Moondream 3 is on our catalogue, and nothing in this article claims we serve either. What we do route is the hosted vision tier — Qwen3-VL-8B, for instance — with provider list price passed through at 0% markup and automatic failover between upstreams. If the self-hosting maths above is what makes you hesitate, that is the option the comparison leaves out, and it is a different purchase rather than a better one.

The licence may decide before the benchmark does

The two licences restrict different things, and which one bites depends on what your company looks like.

LFM Open License v1.0 defines a threshold of $10,000,000 in annual revenue. Below it, commercial use is licensed outright. Above it, section 5 states plainly that commercial use of the work or a derivative is not licensed under the agreement at all. So a startup can build a product on LFM3-Omni-2B for free, and a company past the threshold cannot without a separate deal — regardless of how good the model is.

Moondream Model License 1.0 has no revenue threshold. It permits commercial use at any size, self-hosting, fine-tuning, quantisation and redistribution, and it forbids one specific business: providing a general-purpose hosted model service for the model or a derivative without a separate written commercial licence. Embedding Moondream inside a narrower product is allowed. Reselling general-purpose inference on it is not.

Read together, the two documents are almost complementary. Liquid gates by who you are; Moondream gates by what you sell. A large company planning a hosted inference product is blocked by both, in different ways.

Who should pick which

Pick Moondream 3.1 if the product needs coordinates. Screen automation, robotics, retail shelf audits, counting, anything where the answer is a box, a point or a mask and the next system consumes it programmatically. The 9B-total footprint is the price of that guarantee, and if you are deploying on Jetson-class hardware it is a price you can pay. Plan the hosting migration now rather than in November.

Pick LFM3-Omni-2B if the product involves a person talking to a device, or a device that has to hear a room. A kiosk that answers out loud, a field tool where hands are busy, a camera that has to understand a video with its soundtrack, a single checkpoint replacing an ASR model plus a VLM plus a TTS model. The vision scores are respectable rather than exceptional, the toolchain is a day old, and the speech path is the reason to accept both.

Pick neither, yet, if the decision is load-bearing. Nothing here has been independently evaluated. LFM3-Omni-2B's tooling is moving fast — an Ollama library entry exists at 2.4 GB, and llama.cpp merged support for it on 27 October, the same day the weights appeared — but the Ollama listing currently disagrees with the model card in two visible ways, showing a 125K context and "Text, Image" as the supported inputs against the card's 32K and four modalities. That is a listing that has not caught up, not a model defect, and it is the kind of thing that gets fixed in a week. It is also not something to put in production this week.

What would change this

Two things. An independent evaluation of LFM3-Omni-2B on the audio-visual rows would settle whether the 70.09 on Daily-Omni survives a harness it did not choose, and that is the single number the whole omni case rests on. And a serving story for the speech branch — the model card documents text generation in detail and says much less about how the audio decoder is served at scale — would tell you whether the "one model replaces four" pitch survives contact with a production deployment.

Until then the comparison has a clean shape. Moondream 3 measures the world and hands you numbers. LFM3-Omni-2B listens to it and talks back. The only benchmark they share is a document-reading test, and the omni model wins it by less than four points. That is not a rivalry. It is a reminder that "2B multimodal" describes a memory budget, not a job.