UserRepGen-4B: novidades, hardware necessário, como rodar grátis

UserRepGen-4B turns a chat into a structured user profile. It’s a free, 4-billion-parameter model that writes JSON, not dialogue.

UserRepGen-4B is a fine-tuned version of the Qwen3.5-4B language model. According to Microsoft, it accepts English, multi-turn user-assistant chats and returns a JSON object containing two fields: task and user_behavior. The model’s backbone remains a conditional-generation transformer, but the fine-tune adds a rank-64 LoRA layer trained with online supervised fine-tuning. Its purpose is to create a compact, natural-language representation of a user that can later condition a separate LLM-based simulator.

Unlike earlier UserLM models that generate simulated user turns, UserRepGen-4B produces a description of the user’s task and behavior instead of dialogue. The model does this directly from the source conversation, so it does not require any simulation rollouts or hindsight feedback at inference time. Microsoft says the training used online supervised fine-tuning with hindsight distillation, letting the model learn from comparisons between simulated and original user behavior.

The model card highlights several capabilities. It can capture a user’s communication style, incremental constraint disclosure, feedback patterns, and termination behavior. When fed to an unseen LLM-based simulator, these representations improve conversation-reconstruction fidelity, according to the research. The training sequence length is 9,216 tokens, allowing fairly long multi-turn dialogs to be processed in a single pass.

UserRepGen-4B infers a user profile from just one conversation, so it can over-generalize occasional behavior into a stable rule. The model may include assumptions that are not grounded in the source text, and it does not anonymize the output. Microsoft warns that the generated representations can preserve identifying details and increase the ability to link conversations from the same user, posing a privacy risk.

According to the model card, the primary audience is academic and industry researchers working on LLM-based user simulation, conversation reconstruction, or behavioral diversity studies. Engineers who need a compact, natural-language description of a user to condition a separate simulator will find it useful. The model is free under an MIT license, downloadable from Hugging Face, and runs on modest hardware such as an RTX 4060.

O que você precisa para rodá-lo em casa? Com 4 bits, ele ocupa cerca de 5 gigabytes de memória. A menor máquina doméstica que o comporta por inteiro é RTX 4060. Em RTX 3060 12 GB com 32 gigabytes de RAM, ele escreve cerca de 86 tokens por segundo. Em RTX 4090 com 64 gigabytes de RAM, ele escreve cerca de 208 tokens por segundo.

Ainda não há API grátis para ele. Os pesos são abertos: você pode baixá-los do Hugging Face sob a licença MIT. A Arena ainda não o classificou: adicionamos a pontuação assim que os votos chegarem.

Compare UserRepGen 4B with every free model, and check what your own GPU can run, on FreeLLMArena.