Microsoft releases UserRepGen-4B — a user description generator for simulators

The model microsoft/UserRepGen-4B with the text-generation task has been published on HuggingFace. It is a fine-tuned Qwen3.5-4B that, from a multi-turn dialogue between a user and an assistant, generates a structured user description in JSON: the task separately and the behavior separately. The description is intended for conditioning a separate LLM user simulator — the model itself does not generate replies and does not perform the task. Training went through online supervised fine-tuning with hindsight distillation on comparing simulated and real behavior, but at inference the model outputs the representation directly, without simulation runs.

What it means

The model weighs 4.5 billion parameters, the largest file is 8.5 GB in safetensors. MIT license, running via transformers. There is already one quantized build in GGUF on the site: bartowski/UserRepGen-4B-GGUF, the smallest working one is 2.3 GB in IQ4_XS. To run a 4B model at full precision you need about 9—10 GB of VRAM, the IQ4_XS quant fits in roughly 3—4 GB — that is the level of mid-range local GPUs or CPU inference at acceptable speed. The model is highly specialized: it is needed by those who build user simulators for testing dialogue systems. It is not intended for text or chat generation. You can follow new releases from Microsoft and other developers in the CPU3D news feed.