UserRepGen-4B vs LFM2.5 2.6B Base: The Raw Material and the Finished Part

The quickest way to see what separates microsoft/UserRepGen-4B from LFM2.5-2.6B-Base is to count the formats. Liquid AI's 2.69-billion-parameter checkpoint ships four ways — native Transformers weights, GGUF for llama.cpp, ONNX for cross-platform runtimes, and MLX for Apple Silicon — because it exists to be adapted and deployed by other people. Microsoft's UserRepGen-4B ships one way: a single unquantised bfloat16 safetensors file at 9,078,620,536 bytes, with a card that states plainly that "no optimized or device-specific formats are currently available." One is raw material. The other is a finished part. They are not competitors; they are two consecutive stages of the same job, and putting them side by side is the clearest way to see what actually changes between starting a fine-tune and shipping one.

The two positions in the stack

LFM2.5-2.6B-Base is the pre-trained checkpoint underneath Liquid AI's LFM2.5 family — the one all the post-trained variants are built from. It has deliberately never been instruction-tuned. Its own card recommends it only for "tasks that require heavy fine-tuning, like language-specific (e.g., Japanese) or domain-specific (e.g., medical) assistants, training on proprietary data, or experimenting with novel post-training approaches." Ask it a question and it will continue your text rather than answer it, because answering was never part of its training.

UserRepGen-4B is what the next step looks like. Microsoft took Qwen3.5-4B, added a rank-64 LoRA, trained it with online supervised fine-tuning through hindsight distillation across July and August 2026, and published the result on 14 October with no announcement and no paper. It does exactly one thing — reads a multi-turn conversation and writes a two-field JSON description of the user in it — and it does nothing else.

So the honest framing is not "which is better." It is: one of these is the thing you fine-tune, and the other is what a lab produced by fine-tuning something else.

What the download actually hands you

• Parameters and footprint — LFM2.5-2.6B-Base is 2.69 billion parameters in a hybrid architecture; UserRepGen-4B is a 4B-class model whose published weight file is about 9.1 GB in bfloat16 with no smaller build available.

• Context — 131,072 tokens on the Liquid checkpoint, against a training sequence of 9,216 tokens on the Microsoft one (8,192 prompt, 1,024 completion), with extended use explicitly unsupported.

• Formats — GGUF, ONNX and MLX alongside native weights, versus one safetensors file.

• Licence — the LFM Open License (lfm1.0) on one side, MIT plus an inherited Apache 2.0 on the other. Both permissive enough to build on; neither is the same document.

• Languages — sixteen named languages on the Liquid card, English only on Microsoft's.

• Adoption — LFM2.5-2.6B-Base has been public since 1 August 2026 and sits at roughly 6,282 downloads and 51 likes. UserRepGen-4B, uploaded on 14 October, was at zero of each when this was written.

Only one of those rows is about quality. The rest is about whether the thing fits through the door you have.

Two architectures built for opposite rooms

Liquid describes LFM2.5 as "a new family of hybrid models designed for on-device deployment," and the base checkpoint's shape follows from that. Thirty layers, twenty-two of them double-gated short convolution blocks and eight grouped-query attention, a 128,000-token vocabulary and a 34-trillion-token training budget. Convolution blocks are cheap in memory and predictable in latency; the whole design is aimed at phones, laptops and edge boxes.

UserRepGen-4B inherits the opposite assumption. Its backbone is a standard Qwen3.5 conditional-generation stack — 32 layers, 2,560 hidden size, 16 attention heads with 4 key-value heads, a 248,320-token vocabulary, and a vision tower carried along untouched from the parent checkpoint even though the model is text-only and images are explicitly unsupported. Microsoft reports hosting it with vLLM on H100, A100, RTX A6000 and B200 GPUs. That is a server part list, not an edge one.

A team choosing between them is really choosing which constraint to accept: on-device memory and a runtime you have to fine-tune into shape, or datacenter hardware and a model that already does its one job.

Where the fine-tuning actually moved the number

This is the part of the comparison that a base model cannot answer, by definition. LFM2.5-2.6B-Base publishes no task benchmark, because a base checkpoint has no task — quality numbers for the family belong to the post-trained sibling.

UserRepGen-4B publishes one, and it is the strongest evidence in the whole release. On a 700-conversation fidelity evaluation, judged on a 1-to-10 scale:

• Qwen3.5-4B base, the unadapted backbone — 3.41 best-of-1, 4.42 best-of-3.

• GPT-5.5 prompted to do the same job off the shelf — 3.90 and 4.90.

• UserRepGen-4B — 4.95 and 6.02.

Every figure is Microsoft's own, produced in a loop where GPT-5.4 mini acted as simulator, assistant and judge at once. Nothing has been independently reproduced. But the shape of the result is the point: fine-tuning a 4B base past a frontier generalist at a narrow task is a real, measurable outcome, and it is precisely the outcome LFM2.5-2.6B-Base exists to make possible for somebody else's task. The Microsoft card also shows that the two obvious reinforcement-learning alternatives — OPSD at 4.46/5.60 and GRPO at 3.57/4.59 — both lost to plain supervised fine-tuning, which is a warning about assuming the heavier method wins.

One of these has a privacy page, and it is not the base

A pre-trained base model has no privacy section because it makes no claims about people. UserRepGen-4B has one because it does. Its card reports that over 147 users, a selector picking which of four conversations came from the same person scores 51.02% using the model's behavior descriptions, against 44.22% on the raw conversations and 25% at random — the summary identifies the person better than the source. Microsoft labels it a linkability risk and states that representations are not anonymised data.

That asymmetry is structural, not a defect in either release. The moment a model's output is a description of a human being, the ethical surface changes, and the card has to carry it. Anyone fine-tuning LFM2.5-2.6B-Base toward a similar target should expect to write the same section for themselves.

What the choice of base decides later

There is a second-order consequence to borrowing a base that rarely gets considered at fine-tuning time: it determines what you can call six months later. Microsoft built on Qwen3.5-4B, and that family is served commercially — Qwen3.5-Flash over a million-token context at $0.10 per million input tokens and $0.40 per million output, Qwen3.5-27B at $0.086 and $0.688, Qwen3.5-35B-A3B at $0.057 and $0.459, all reachable through one OrcaRouter key at the provider's list price with nothing added, and all able to sit behind automatic failover so a research checkpoint that only exists on your own disk does not become a single point of failure in a pipeline. We do not carry UserRepGen-4B, and we do not carry LFM2.5-2.6B-Base either — both are downloads, and the small-model fine-tune you produce from the Liquid checkpoint will be yours to host. But the assistant side of the loop, or the judge, or the fallback path, can be an API call on a key you already have. That is a design decision you make at the start, when you pick which base to adapt.

Which one you actually want

If you have a task and no model for it, LFM2.5-2.6B-Base is the right starting point: small enough to fine-tune on modest hardware, hybrid enough to run on a device, and shipped in every format you will need once the training is done. If you have a user-simulation pipeline and need a component that turns a finished conversation into a portable behavioural description, UserRepGen-4B already exists, and the fact that it is one file in one format is a limitation rather than a reason to redo the work. The thing to watch is whether Microsoft follows it with the paper its card still promises. For Liquid the open question is different and older: which teams pick up the base and what they make from it.

Base models exist to be adapted, and adaptation is where things break: the documented AI incidents and agent failures around 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.