UserRepGen-4B vs Intern-Decision-4B: Two Teams Fine-Tuned the Same 4B Base and Agreed on Nothing

Two quiet releases, five weeks apart, branching from the same checkpoint. microsoft/UserRepGen-4B went up on Hugging Face on 14 October 2026 as a 9.1 GB bfloat16 repository with no announcement; internlm/Intern-Decision-4B appeared on 26 September 2026 at 05:36:37 UTC inside a forty-second window that also produced a 0.8B and a 2B, equally unannounced. Both are LoRA fine-tunes of Qwen3.5-4B, both are text-only despite carrying a vision tower they never use, both return a rigid structure rather than an answer — and they could not disagree more about what the model is for. Intern-Decision-4B never writes a token. UserRepGen-4B writes nothing but tokens. Everything else in this comparison follows from that one split.

One is a scorer, the other is a writer

InternLM's card is unusually blunt about its own mechanics: "This API performs structured candidate scoring. It does not call generate() or sample free-form text." You hand it a state, a schema of named questions, and a set of candidate answers. It maps the options onto single-token symbols, runs one causal forward pass, reads the logits at the position immediately before each placeholder, applies a fitted temperature of 1.99241824, and returns a calibrated probability per option. One pass. No sampling, no decoding loop, no text.

UserRepGen-4B inverts it. You hand it a completed conversation. It emits up to 1,024 tokens of prose across two JSON fields — a task description and a user_behavior description — and the card is equally blunt in the other direction: the model "generates representations, not simulated user turns," it does not execute the user's task, and it is not the simulator.

Put the two side by side and the engineering consequences fall out immediately.

• What the output is worth — Intern-Decision-4B returns a number you can threshold, average and audit. UserRepGen-4B returns a paragraph whose quality only an LLM judge can score, which is exactly how Microsoft evaluated it.

• Cost per call — one forward pass over a fixed prompt versus up to 1,024 generated tokens. That is the difference between roughly 44 milliseconds and a full decoding run, and it is why the decision model can sit inside a loop that the representation model cannot.

• Where the meaning lives — InternLM decided the answer should already be enumerable in the prompt, so the model only has to rank. Microsoft decided the answer is latent in a conversation, so the model has to write it. Ranking is bounded and verifiable. Writing is open-ended and needs a judge.

The input ceilings, and what they tell you about the bets

Both cards publish a hard limit, and comparing them is the fastest way to see how differently the two teams scoped the same base model.

Intern-Decision-4B refuses work above 8,192 tokens. That is a refusal, not a truncation — a design choice that says the model is for decisions compact enough to state completely in a prompt.

UserRepGen-4B is trained on a 9,216-token sequence: up to 8,192 prompt tokens covering the conversation, plus 1,024 completion tokens for the representation. The card adds that extended-context use beyond that length "is not supported by this release," and that long-context performance was never separately benchmarked. So the same 8,192-token ceiling shows up on both sides, arrived at independently, from two teams who presumably had no visibility into each other's work.

That convergence is the useful finding. A 4B fine-tune on this base, in the middle of 2026, is a component sized for a single well-formed prompt — not a long-context engine. Both releases are honest about it. Neither hides it behind a bigger advertised window.

Both published a calibration story, and they are not the same story

This is where the sourcing gets delicate, because the two evaluations are not commensurable and neither has been reproduced outside its own lab.

Intern-Decision-4B publishes a seven-suite benchmark where it scores 90.02 on average — the best row in its own table, ahead of Intern-Decision-2B at 84.68 and Intern-Decision-0.8B at 79.38, all measured on InternLM's harnesses. It also publishes calibration: a Brier score of 0.347 and an expected calibration error of 0.065. A separate 96-case pilot using exact reference distributions moves from 0.628 Brier and 0.213 ECE to 0.550 and 0.089 after its fitted temperature is applied. Every one of those numbers is the vendor's own, and nobody outside InternLM has checked them.

UserRepGen-4B publishes a fidelity table on a 1-to-10 judge scale — 4.95 best-of-1 and 6.02 best-of-3, against 3.90 and 4.90 for GPT-5.5 used off the shelf and 3.41 and 4.42 for the untuned base. Every one of those numbers is Microsoft's own, produced with GPT-5.4 mini acting as simulator, assistant and judge simultaneously, and nobody outside Microsoft has checked them.

The asymmetry that matters is not the scores — they measure different things on different scales. It is that Intern-Decision-4B publishes a proper scoring rule against known reference distributions, which is falsifiable by construction, while UserRepGen-4B publishes an LLM-judge score, which is falsifiable only by running a different judge. InternLM built a model that can be audited with arithmetic. Microsoft built one that has to be audited with another model. Given that both are claiming to have beaten baselines, that difference in auditability is a real part of what you are choosing between.

The privacy finding only one of them prints

Intern-Decision-4B's card contains no privacy section of consequence. UserRepGen-4B's contains a number that should give any team pause: over 147 users, a selector asked to pick which of four conversations came from the same person hits 51.02% using the model's behavior descriptions, against 44.22% using the full raw conversations and 25% at random. The compressed representation identifies the person better than the material it was derived from.

Microsoft labels it a linkability risk and says outright that representations are not anonymised data. That is the correct handling, and it is a caution that applies to the whole class of models — including the decision model, whose own inputs are conversations about people's situations. The difference is only that one card says so and the other does not, which is not the same as one model being safe.

Neither is an API, but the base they share is

Both of these are downloads, not endpoints. Intern-Decision-4B ships as a Python DecisionEngine class loaded once and reused, its demo Space answers 401, and its GitHub link 404s; UserRepGen-4B ships a single unquantised safetensors file and an example.py. OrcaRouter carries neither. If you are saying either one is available through a hosted API today, you are describing a plan rather than a fact.

What is genuinely callable is the checkpoint both of them branched from. Qwen3.5-4B's family is on our catalogue — Qwen3.5-Flash over a million-token context at $0.10 per million input tokens and $0.40 per million output, Qwen3.5-35B-A3B at $0.057 and $0.459, Qwen3.5-27B at $0.086 and $0.688 — all on the same key, all at the provider's list price with nothing added. That is useful for the part of the pipeline that is not these two models: the assistant side of a simulation, or the reasoning step around a decision call, can run through one endpoint while you evaluate the fine-tune locally.

How to choose

If your problem is a bounded question with enumerable answers and you need a calibrated number ten thousand times a second, Intern-Decision-4B is the only model here that does it, and the fact that it never samples is the feature rather than the limitation. If your problem is that you want a simulator that behaves like a real user, UserRepGen-4B is the only model here trained for it, and the size of its reported gain over GPT-5.5 is the reason to care. Nobody should pick either on the basis of the other one's numbers, because they never measured the same thing. What would settle the comparison is a single piece of external scrutiny on either card — and for two releases that shipped in total silence, weeks apart, neither has had any.

Two quiet releases are also a reminder of how much of this class of model ships without documentation, and the failures that follow are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.