UserRepGen-4B vs Gemma 4 12B: A 4B Narrow Specialist Against a 12B Generalist, and Why the Sizes Lie

Placed side by side, microsoft/UserRepGen-4B and Gemma 4 12B look like a mismatch decided before it starts. G​oogle's 11.95-billion-parameter model reads text, images and audio, holds 256K tokens of context, carries a technical report and a launch blog, and has been downloaded around 79,000 times since May 2026. Microsoft's model is a rank-64 LoRA on Qwen3.5-4B, uploaded on 14 October 2026 with no announcement, and it does exactly one thing: it reads a finished user-assistant conversation and writes back a two-field JSON description of the person in it. It cannot hold a conversation, answer a question, or see an image. Weighing these two on capability is not a comparison, it is a category error — and the category error is the interesting part, because the smaller model is the one that beats a frontier generalist at the job it was built for.

What each one is actually for

The cleanest way to hold both in your head is that Gemma 4 12B is an assistant and UserRepGen-4B is a measurement instrument.

• Gemma 4 12B — the encoder-free multimodal member of the Gemma 4 family. Text, image and audio go in; text comes out. G​oogle's own card describes it as "Unified" because it drops the dedicated vision and audio encoders the rest of the family uses, projecting raw image patches and waveforms straight into the decoder's embedding space through lightweight linear layers. 48 layers, a 1,024-token sliding window, a 262K vocabulary, 256K-token context. Apache 2.0, instruction-tuned variants published alongside the base.

• UserRepGen-4B — a representation generator. Input: an authorised multi-turn conversation, up to 8,192 tokens. Output: a JSON object with a task field and a user_behavior field, capped at 1,024 tokens, reviewed by a human and then handed to a separate LLM that plays the user in a simulation. MIT on Microsoft's contribution, Apache 2.0 inherited from the Q​wen backbone. Text only, despite the checkpoint retaining a vision tower it never uses.

Gemma 4 12B can be asked to do what UserRepGen-4B does — read a conversation and describe the user in it. That is precisely the honest way into this matchup, because Microsoft's card measures a general model doing exactly that, and it loses.

The only real number that connects them

There is no published head-to-head between these two, and inventing one would be the easy failure of a page like this. What the UserRepGen-4B card does publish is a 700-conversation fidelity evaluation on a 1-to-10 scale, where every generator is judged by identical calls. In that table, GPT-5.5 used off the shelf as a representation generator scores 3.90 best-of-1 and 4.90 best-of-3. UserRepGen-4B scores 4.95 and 6.02. The trained 4B specialist beats a frontier generalist that is orders of magnitude larger at this specific job.

Gemma 4 12B is not in that table, so nobody knows where it would land. But the baseline it would be measured against makes the shape of the answer clear: a general model asked to describe a user produces something plausible and generic, and the training method in UserRepGen-4B exists specifically to push past that. A 12B will very likely beat the 4B on general reasoning and be beaten by it on this narrow task.

Which raises the obvious question of why you would not simply prompt Gemma 4 12B to do the same thing and keep one model on your disk. Two reasons, and the second is the one that matters.

The first is cost at volume. Generating a representation per conversation is a batch job, and 12B parameters cost roughly three times what 4B parameters cost for the same pass. The second is that the fidelity gap is not a tuning problem — the card's own ablations show plain supervised fine-tuning at 4.64/5.68 while both OPSD and GRPO, the two reinforcement-learning methods a competent team would reach for first, land below it at 4.46/5.60 and 3.57/4.59. Getting a general model to this behaviour is not a matter of a better prompt or a heavier RL run. It is what the training loop was built for.

Both are downloadable, and only one of them is deployable

This is where the 4B-versus-12B framing falls apart completely, because the two releases are at opposite ends of the packaging spectrum.

• Gemma 4 12B — a released product with a working ecosystem. Transformers support, an instruction-tuned variant, quantised and edge formats published across the family, a launch blog, a technical report on arXiv, and a documented path from download to serving.

• UserRepGen-4B — a single bfloat16 safetensors file at 9,078,620,536 bytes. No GGUF, no ONNX, no MLX, no quantisation of any kind. The card says so directly: "No optimized or device-specific formats are currently available."

That gap is not incidental. Gemma 4 12B was built to be run by other people, and eleven and a half billion parameters ship as a product. UserRepGen-4B was built to produce a result for a paper, and its card still says the citation is "pending confirmation" and refers to authenticating "while the repository is private" — a line left over from a staged upload that has not been updated. One is finished; the other is finished enough.

If what you actually wanted was a Gemma 4 through an API

Worth saying plainly, because it is where most readers of a Gemma comparison actually live: OrcaRouter does not serve Gemma 4 12B — there is no model page for it on our catalogue — and it does not serve UserRepGen-4B either. Two other members of the same Gemma 4 family are on the catalogue, though, and both quote the provider's list price with nothing added: Gemma 4 26B A4B, the 25.2B mixture-of-experts build with 3.8B active parameters over a 262,144-token context, at $0.06 per million input tokens and $0.33 per million output; and Gemma 4 31B, the dense 30.7B multimodal build, at $0.13 and $0.38. If you wanted the 12B specifically for local work, the download is the answer and no API changes that.

For the simulation side the honest position is the same. UserRepGen-4B is a checkpoint and a prompt file — you run it yourself, in your own environment, against conversations you are authorised to process. It would be convenient if that were a call to a hosted endpoint. It is not, today, on ours or anywhere else.

Which one to download

If you need a small multimodal assistant that reads documents, hears audio and answers questions on a laptop, Gemma 4 12B is the one, and the comparison to UserRepGen-4B is not even close enough to be relevant. If you are building a user simulator and your current approach is to prompt a general model to describe the user, UserRepGen-4B is the only thing here that was trained for that, and the evidence — vendor-reported, single-lab, judged by GPT-5.4 mini throughout — says the training beats the prompting. What would change the picture is the paper, or one independent evaluation of that 4.95 from outside Microsoft. Neither exists yet, and both are worth waiting for before anyone replaces a working pipeline with it.

Comparisons like this one are also where the failure modes surface: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.