Quick answer
The comparison in the launch post is not like-for-like. Liquid AI benchmarked its 1.9B vision model against Gemma 4 12B. Google also ships E2B, a 2B-effective model for the same ultra-mobile job
No valid head-to-head score exists. Liquid AI reports an average over its own 62 evaluations; Google publishes a different suite. They cannot be compared directly
Licence decides it for most teams. Gemma 4 12B states Apache 2.0 on its model card. The LFM3 launch post names no licence at all
Footprint is close but not measured the same way. LFM3 text at about 1GB at 4-bit; Gemma 4 E2B at about 1.1GB as BF16 static weights
Gemma 4 scales, LFM3 specialises. Gemma 4 runs 2B to 31B under one family; LFM3 adds decision models Gemma has no answer to
Liquid AI released six LFM3 models on 27 October, all around 2B parameters, and the number everyone quoted was the vision model averaging 71.4 across 62 evaluations against 71.7 for Gemma 4 12B. A 1.9B model landing three tenths of a point behind a 12B is a genuinely good result. It is also not the comparison a team picking a model has to make.
The comparison the launch post skipped
Google does not only ship a 12B. The Gemma 4 family runs E2B at 2B effective, E4B at 4B effective, the 12B, a 26B mixture-of-experts with about 4B active per token, and a 31B dense model. Google describes E2B and E4B as built for "ultra-mobile, edge, and browser deployment" and puts E2B at about 1.1GB on mobile.
That is the same socket LFM3 is competing for. Benchmarking a 2B against a 12B that needs 26.7GB of GPU or TPU memory flatters the small model by comparing it to something that was never going to run on the phone anyway.
| Liquid AI LFM3 | Google Gemma 4 |
|---|
| Released | 27 October 2026 | 2026, four to five sizes |
| Sizes | Six models, all around 2B, branching from a 1.7B base | E2B (2B effective), E4B (4B effective), 12B, 26B A4B (about 4B active), 31B dense |
| Modalities | Text, vision, audio, omni, plus two decision models | Text, image, video and audio. Native audio on E2B, E4B and 12B |
| Context | Not stated in the launch post | 128K on the small models, 256K on the medium. The 12B card states 256K |
| Mobile footprint | Text model about 1GB at 4-bit | E2B about 1.1GB and E4B about 2.5GB on mobile, BF16 static weights via LiteRT-LM |
| Named hardware | Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, Qualcomm Hexagon NPU | Ultra-mobile, edge and browser for E2B and E4B; 26.7GB GPU or TPU for the 12B |
| Licence | Not named in the launch post. Weights published on Hugging Face | Apache 2.0 on the official 12B model card. Google’s overview page refers to a Gemma 4 licence permitting responsible commercial use |
| Decision models | d1-2B and d1-Omni-2B, built on the vision and omni models | No equivalent |
Why we are not scoring this
Liquid AI reports an average of 71.4 across 62 evaluations it selected. Google publishes MMLU Pro, GPQA Diamond, AIME 2026, LiveCodeBench and others on its model card. These are different benchmark sets measured by different parties, and averaging across them to declare a winner would produce a number with no meaning.
| Model | Source of the number | Reported |
|---|
| LFM3-VL-2B (1.9B) | Liquid AI, own selection of 62 evaluations | Average 71.4, against 71.7 for Gemma 4 12B and 73.4 for GPT-5.6 Luna |
| LFM3-1.7B | Liquid AI | Average 49.37 against 31.55 for LFM2.5-2.6B. IFEval 96.66, ToolSandbox 80.36, BFCL V4 71.39 |
| LFM3-Audio-2B | Liquid AI | English ASR word error rate 6.2, VoiceBench 73.1, MMAR 50.6 |
| Gemma 4 12B | Google, official model card | MMLU Pro 77.2%, GPQA Diamond 78.8%, AIME 2026 77.5%, LiveCodeBench v6 72.0%, MMMU Pro 69.1% |
| Gemma 4 E2B | Google | Not published in the overview page we read |
Read the provenance column, not the scores. Every LFM3 figure here comes from Liquid AI describing its own models, including the Gemma 4 comparison point. Every Gemma 4 figure comes from Google describing its own. Neither party benchmarked the other under the same conditions, and nobody independent has benchmarked both. If you need a decision, run both on your own task.
The licence is the part that actually decides it
Google’s official Hugging Face model card for Gemma 4 12B states Apache 2.0. That is about as unambiguous as open weights get: use it commercially, modify it, ship it. One wrinkle worth knowing is that Google’s own model overview page instead refers to a "Gemma 4 license" that permits responsible commercial use, so the two Google surfaces are not phrased identically. Read the card for the model you actually pull.
Liquid AI calls LFM3 open-weight and publishes to Hugging Face, the Playground and its own API. The launch post does not name a licence. That is not an accusation of anything - the model cards may well carry a permissive one - but a company deciding what to embed in a shipped product cannot act on "open-weight" alone, and this is exactly the distinction that produces an awkward conversation with legal six months in.
Which to choose
Take Gemma 4 if the model goes into something you sell. Apache 2.0 on the model card settles the question that matters most, the family spans 2B to 31B so you can move up a size without changing vendor or rewriting your integration, and E2B targets the same phones LFM3 does. The 256K context on the 12B is also well ahead of anything stated for LFM3.
Take LFM3 if your problem is one of the shapes it was built for. The decision models, d1-2B and d1-Omni-2B, have no Gemma equivalent at all, and the audio model is a first-class member of the family rather than a capability bolted to a general model. The explicitly named hardware list, Raspberry Pi 5 included, is also more concrete than Google’s category descriptions if you are deploying to something unusual.
Take neither yet if you are choosing on the 71.4 figure. It is one vendor’s average over its own evaluation selection against a model in a different weight class. Spend an afternoon running both on fifty of your own examples and the answer will be better than anything in this table.
FAQ
Is LFM3 actually better than Gemma 4?
No honest head-to-head exists. Liquid AI reports an average across its own selection of 62 evaluations; Google publishes a different benchmark set on its model card. Averaging different suites and calling the result a winner would be meaningless. What you can compare is size, modality coverage, footprint, context and licence, which is what the table above does.
Which one can I use in a commercial product?
Gemma 4, on the evidence available: the official 12B model card states Apache 2.0. Liquid AI calls LFM3 open-weight and publishes on Hugging Face, but the launch post names no licence, so check each model card before you build. Open weights and a permissive licence are different things.
Is the 1GB versus 1.1GB difference real?
Not as stated. Liquid AI quotes about 1GB for its text model at 4-bit precision; Google quotes about 1.1GB for E2B as BF16 static weights via LiteRT-LM. Different precisions and different accounting. Quantise Gemma 4 E2B to 4-bit and the comparison changes. Measure on your own target device.
Why compare against E2B rather than the 12B?
Because E2B is the model built for the job LFM3 is built for - ultra-mobile, edge and browser deployment. A 12B needing 26.7GB of GPU or TPU memory is not competing for the same socket as a model that fits in about a gigabyte on a phone.
Does either run on a Raspberry Pi?
Liquid AI names the Raspberry Pi 5 explicitly. Google describes E2B and E4B as ultra-mobile, edge and browser models without naming that board, so treat Pi support as unverified on the Gemma side until you test it.
What does LFM3 have that Gemma 4 does not?
Decision models. d1-2B and d1-Omni-2B are built on the vision and omni models and reported against Decision Bench at 85.2 and JevBench at 89.0. Gemma 4 ships no equivalent, so if that is the shape of your problem the comparison is moot.
Sources
Liquid AI on LFM3, 27 Oct 2026 ·
Google, Gemma 4 model overview ·
Official Gemma 4 12B model card ·
Our 29 October news roundup