LFM3-Omni-2B vs InternLumina U2: Two Models, Two Meanings of Omni

LFM3-Omni-2B and InternLumina U2 are both filed under "omni" and they are not in the same business. Liquid AI's 2.17-billion-parameter model takes text, images, video and audio in and gives back text or 24-kilohertz speech. Shanghai AI Laboratory's InternLumina U2 takes text, images, video and 3D in and gives back text or pictures. One is a mouth. The other is a hand. Almost every comparison you will find between them spends its length on benchmarks that only one side can even sit, which is why the useful question is narrower and cheaper to answer: which half of "omni" is your product missing?

The two releases also arrived in opposite shapes, and that matters more than it usually does. Liquid AI announced the LFM3 family on 27 October 2026 in a single post covering six 2B-class models, and LFM3-Omni-2B was one line of that post rather than a launch of its own — its weights had already been sitting in a public repository since 21 October. InternLumina U2 went the other way: the inference code and a project page appeared on 1 September, the weights followed later in hardware-specific folders, and no announcement ever came. Neither model has a number that a third party has reproduced. Every figure below is the vendor describing itself.

The refusals tell you more than the specs

Spec lists are written to make models look compatible. Start instead with what each one physically cannot accept, because the two lists do not intersect at all.

LFM3-Omni-2B cannot draw. There is a vision encoder in the stack, but it only reads; the output path is text and speech. Ask it to edit a photo and the honest answer is that the request has no representation in the model.

InternLumina U2 cannot speak. It has no audio encoder, no codec tokenizer, no speech head. Hand it a WAV file and there is nothing in the architecture to receive it. Its "omni" is a bet that understanding and generation belong in one visual backbone; its output vocabulary is text and image tokens, not waveforms.

So the first question is not which model scores higher. It is whether your workload needs a synthesized voice or an edited image, because exactly one of those is available on each side and neither model will grow the other capability by being fine-tuned. That is a property of the tokenizer, not of the training data.

One dictates a sentence, the other refines it

The architectures differ at a level deeper than parameter count, and the difference explains the capability split.

LFM3-Omni-2B is autoregressive. A 1.7-billion-parameter LFM3 backbone reads tokens from an LFM3 vision encoder and an LFM3 causal audio encoder, and emits text tokens or LiquidCodec speech tokens left to right. Video is sampled at 2 frames per second and each 2×2 image patch collapses to a single language token, which is how a 2.17B model keeps a 32K window usable on a phone. Speech comes out at 24 kHz, audio goes in at 16 kHz mono.

InternLumina U2 is a diffusion language model, and that is the part worth slowing down for. It does not emit one token at a time; it starts from noise and refines a whole sequence in parallel, using a masked-diffusion objective. Its visual representation is fully discrete — eight codebooks of AToken visual tokens per position, concatenated on the input side and predicted codebook by codebook on the output side. Generating an image is then the same operation as answering a question: denoise a sequence of visual tokens. That is a genuinely different bet from the one every other open multimodal model is making, and it is why the model can produce a picture at all.

The cost of that bet is tooling. InternLumina is not a transformers pipeline you call in three lines. The card requires the AToken tokenizer with FlashAttention, a custom block-diffusion generate API, and Blender 4.5 if you want the 3D output rendered. LFM3-Omni-2B, by comparison, ships as ordinary BF16 weights with a documented fine-tuning path and an Ollama build already circulating — one whose listing quotes a 125K context against the 32K in the model card, a discrepancy nobody has explained yet.

The specs that decide whether it fits

• Parameters — LFM3-Omni-2B 2.17B dense; InternLumina U2 16B total, 1B active in a sparse mixture of experts.

• Output — text and 24 kHz speech, including interleaved text-and-speech, versus text and generated or edited images.

• Input — text, image, video with or without audio, and 16 kHz audio, versus text, image, video and 3D.

• Context — 32K tokens covering input and output, against a context length the InternLumina card still does not state.

• Languages — 16 text languages and 19 speech locales for LFM3-Omni-2B; InternLumina does not publish a count.

• Licence — LFM Open License v1.0, a revenue-threshold licence, against Apache 2.0 with carve-outs for the bundled AToken and VeOmni code.

• Hardware — phones, laptops and edge boards on one side; on the other, checkpoints split into an ascend/ tree trained on Huawei Ascend NPUs and an nvidia/ tree, roughly 33.8 GB per tree.

The two benchmarks they actually share

There are exactly two video numbers in both cards, and they disagree about which model is better, which is the most informative thing on this page.

On MVBench, InternLumina U2 reports 61.8 and LFM3-Omni-2B reports 58.03 — a 3.8-point edge to the diffusion model. On Video-MME without subtitles, LFM3-Omni-2B reports 57.70 and InternLumina U2 reports 56.7 — a 1-point edge the other way. Both sets are vendor-run, on different harnesses, with different prompts, and neither lab has published the eval code. A 1-point gap under those conditions is noise; a 3.8-point gap is probably real but says little about your footage.

Everything else is unshared. InternLumina's card leads with CharXiv-DQ 85.9 and HallusionBench 64.6, document and hallucination tests that LFM3-Omni-2B never sat. LFM3-Omni-2B's card leads with DocVQA 92.47, OCRBench 806 and MMMU 42.44, plus a speech column — English ASR word error rate 6.36, multilingual 6.72, TTS UTMOSv2 3.80 — that has no counterpart on the other side. Its own card labels the numbers as measured by Liquid AI with each model's own chat template and greedy decoding, which is the standard caveat and a real one.

InternLumina is more candid still: its card calls the results "preliminary, partial results" and defers the comparison tables. That is not false modesty. A 16B-A1B diffusion model generating images at DPG-Bench 87.79 and GenEval 0.89 is a serious result, and it is also the kind of result that changes when someone reruns it on a different sampler.

Who should pick which

If your product needs to talk back — a voice assistant, a hands-free interface, an accessibility tool, a call summariser that hands the caller a spoken answer — LFM3-Omni-2B is the only one of these two that can finish the job. It is also the only one that runs on the device, which for speech matters twice: audio is the most privacy-sensitive modality most apps touch, and shipping it to a server is a decision you have to justify.

If your product needs to produce or edit an image, InternLumina U2 is the only one that can start. The trade is that you are now maintaining a diffusion inference stack on Ascend-first hardware, with an NVIDIA tree that has only recently become usable and a 3D path that wants a Blender install. That is a research deployment, not an app-store one.

What almost nobody admits is that real products need both, and these two models cannot compose with each other. A request that needs a generated diagram and a spoken explanation has to cross two runtimes, two licences and two hardware profiles. That seam is where most teams eventually put a router rather than a bigger model. OrcaRouter is the hosted side of that arrangement — one API across 200+ models at 0% markup on the provider's list price, with automatic failover when a provider drops — and neither LFM3-Omni-2B nor InternLumina U2 is on that catalogue today. The point is not that either belongs there. The point is that the moment you split the work in two, the second half needs somewhere to run.

What would change the answer

One independent evaluation of either model would do more than another vendor table. For InternLumina U2 the number to watch is generation quality under a different sampler, because diffusion outputs move when the sampler moves and no outside party has tried. For LFM3-Omni-2B it is whether the speech holds up outside Liquid's own prompts — specifically, whether the 3.80 UTMOSv2 survives contact with a listener who is not grading their own work.

Until then, the decision rests on something rarer than a benchmark: a capability list. LFM3-Omni-2B speaks and cannot draw. InternLumina U2 draws and cannot hear. For most teams that settles it before the first eval runs.