UserRepGen-4B Text Generation Model: 4GB+ VRAM, GGUF Builds

- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. How It Differs from Similar Models
- 7. Hardware Requirements
- 8. Can You Run It Locally?
- 9. Recent Models in the Same Size Class
- 10. How to Get It
- 11. Quantized and Converted Variants
- 12. Related Articles
- 13. What to Read Next
- 14. Sources
- 15. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | microsoft/UserRepGen-4B |
| Family guide | Qwen3.5 guide (8 articles) |
| Publisher guide | Microsoft: models and licenses |
| Published | 2026-10-15 |
| License | mit |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Microsoft has released “UserRepGen-4B", a model that generates structured data expressing a user’s “tasks" and “behavior patterns" in natural language from multi-turn user-assistant conversations. It is a fine-tuned model based on Qwen3.5-4B, specialized in creating inputs to be passed to another LLM-based “user simulator".
While other Microsoft models such as UserLM-8b and UserLM-2.0 generate the user’s actual utterances themselves, UserRepGen-4B is characterized by generating only the “representation" and not performing the user simulation itself. It is online supervised fine-tuned using conversation simulation and feedback-based learning (hindsight distillation), but at inference time, it does not require simulation execution or feedback, generating representations directly from the original conversation.
Specifications
- Model type: 4B-class language model for user representation generation (as stated in the model card)
- Architecture: Qwen3_5ForConditionalGeneration (backbone is Qwen3.5-4B)
- Adaptation method: Online supervised fine-tuning (OSFT) with Rank-64 LoRA
- Supported languages: English
- Input: Text only, multi-turn user-assistant conversations
- Output: JSON-formatted text with two fields:
taskanduser_behavior - Training sequence length: 9,216 tokens (maximum 8,192 tokens for prompts + maximum 1,024 tokens for generation)
Performance
The model card includes a table comparing “conversation reproduction fidelity" when simulating users using the generated representations. The scores are under two conditions: Best-of-1 and Best-of-3.
| Representation Generation Method / Training Method | Best-of-1 | Best-of-3 |
|---|---|---|
| Qwen3.5-4B base model | 3.41 | 4.42 |
| GPT-5.5 | 3.90 | 4.90 |
| Supervised fine-tuning (SFT) | 4.64 | 5.68 |
| OPSD | 4.46 | 5.60 |
| GRPO | 3.57 | 4.59 |
| UserRepGen-4B (OSFT) | 4.95 | 6.02 |
This table shows results independently measured by Microsoft’s release team. UserRepGen-4B (trained with the OSFT method) records the highest score among the compared items under both Best-of-1 and Best-of-3 conditions, demonstrating that it creates representations allowing a simulator to faithfully reproduce the original conversation better than when using the base Qwen3.5-4B or the large model GPT-5.5 to generate representations. Compared to other training methods such as general SFT, OPSD, and GRPO, OSFT also achieves the highest values. Note that because the model card’s explanation does not specify what standard benchmark these Best-of-1 and Best-of-3 metrics actually measure, only the relative magnitudes of the scores are presented here. Comparative data for other domains (such as general knowledge or coding) is not listed in the model card.
Strengths and Use Cases
UserRepGen-4B is not a general-purpose dialogue model, but a model specialized for research purposes such as “LLM-based user simulation". The direct use case specified by the model card is converting consented conversation data into compact natural language representations that can be passed to another user simulator, intended to support research on conversation reproduction, simulator evaluation, interactive learning and evaluation of AI assistants, and behavioral diversity in simulation populations. The tags also include user-representation, user-simulation, simulation, and multi-turn, reflecting its specialized use case.
The most suitable targets are English multi-turn user-assistant conversations similar to the 7 datasets used for evaluation. The model card explicitly states that languages other than English, short or single-turn exchanges, non-text modalities, and use cases involving the identification of users’ real-world identities or critical decision-making such as employment, healthcare, and finance are outside the scope of evaluation. Although the checkpoint itself inherits the vision-compatible architecture of Qwen3.5, fine-tuning and evaluation target text input only, and image/video inputs are not supported.
How It Differs from Similar Models
Previously on this site, we introduced “Tev1-4B-experimental" A 4B model specialized in structured data decision-making: Required VRAM 4GB~, which is also based on Qwen3.5-4B. While Tev1-4B-experimental is a classification and decision-making oriented model that returns a single choice label from structured inputs such as “state", “question", and “choices", UserRepGen-4B differs significantly in that it generates the user’s “task" and “behavior pattern" as natural language descriptions from conversation history. Even though they use the same base model, they target completely different directions: one outputs a single selected choice, while the other generates profile-like descriptions.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 4.5B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | Q6_K_S | 3.3GB | 4.0GB |
| 8GB (RTX 4060 / 3060 Ti, etc.) | Q8_0 | 4.2GB | 5.0GB |
| 12GB (RTX 4070 / 3060 12GB, etc.) | BF16 | 7.8GB | 9.4GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build bartowski/UserRepGen-4B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but bartowski/UserRepGen-4B-GGUF provides a GGUF build you can use.
License — mit (Commercial use allowed): Permits commercial use, modification and redistribution, provided the copyright notice and license text are retained.
Compression: the Q8_0 build measures 7.95 bits per weight — about 50% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| IFM/Jais-3-7B-GGUF | 9.0B | 8GB | apache-2.0 | Jais-3-7B-GGUF Arabic and English 7B Reasoning Model: 8GB+ VRAM (2026-11-17) |
| IFM/Jais-3-3.7B-GGUF | 5.1B | 4GB | apache-2.0 | Jais-3-3.7B-GGUF 512K Context Language Model: 4GB+ VRAM (2026-11-17) |
| ibm-granite/granite-switch-4.1-8b | 8.8B | 24GB | apache-2.0 | IBM Releases Granite Switch 4.1: Modular LLMs with aLoRA and Speech (2026-11-17) |
| ggml-org/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF | 8.0B | 12GB | apache-2.0 | NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF Text Generation Model: 12GB+ VRAM (2026-11-17) |
| ByteDance-Seed/Cola2 | 8.8B | 48GB | — | ByteDance-Seed Releases Cola2 Language Model (2026-11-17) |
How to Get It
Distributed in safetensors format, it can be loaded with Hugging Face Transformers’ AutoModelForImageTextToText. The model card reports that vLLM was used to host the inference environment during research on NVIDIA H100, A100, RTX A6000, and B200, and it is expected to work with either Transformers or vLLM. As an example of how to obtain it, the model card shows how to download the included sample script.
hf download microsoft/UserRepGen-4B example.py --local-dir userrepgen-example
python userrepgen-example/example.py --device cuda
For environments without a GPU, --device cpus can be specified, and the model card notes that loading and generation have been confirmed on CPU on Linux (as float32 is used, this is positioned as a functionality check rather than performance measurement). The license is MIT, and agreeing to the Hugging Face terms of use (gated) is not required at this time.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | bartowski | GGUF (imatrix) | bartowski/UserRepGen-4B-GGUF | Q6_K_S 4.0GB (fits in 4GB VRAM) |
| 2026-11-17 | mradermacher | GGUF | mradermacher/UserRepGen-4B-GGUF | Q6_K 3.9GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in bartowski/UserRepGen-4B-GGUF: IQ2_M 1.5GB / Q2_K 1.6GB / IQ3_XXS 1.8GB / Q3_K_S 1.8GB / IQ3_XS 1.8GB / Q3_K_M 1.9GB / Q3_K_L 2.0GB / IQ3_M 2.1GB / IQ4_XS 2.3GB / Q4_K_S 2.4GB / Q4_0 2.5GB / IQ4_NL 2.5GB / Q4_K_M 2.5GB / Q4_1 2.7GB / Q4_K_L 2.7GB / Q5_K_S 2.9GB / Q5_K_M 3.1GB / Q6_K_S 3.3GB / Q6_K 3.4GB / Q6_K_L 3.6GB / Q8_0 4.2GB / BF16 7.8GB
- Available builds in mradermacher/UserRepGen-4B-GGUF: Q2_K 1.8GB / Q3_K_S 1.9GB / Q3_K_M 2.1GB / Q3_K_L 2.3GB / IQ4_XS 2.4GB / Q4_K_S 2.4GB / Q4_K_M 2.5GB / Q5_K_S 2.8GB / Q5_K_M 2.9GB / Q6_K 3.2GB / Q8_0 4.2GB / F16 7.8GB
In addition, 4 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Qwen3.5-4B-GGUF Vision-Language Model: 4GB+ VRAM
- ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM
- Tev1-4B-experimental Text Generation Model: 4GB+ VRAM, GGUF Builds
- llama.cpp Speculative Decoding Test GGUF Models Released
What to Read Next
- Find models by VRAM (This model runs from the 4GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → Qwen3.5 family overview (8 articles, 10 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q6_K_S, Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → Microsoft: models, licenses and articles
Sources
- https://huggingface.co/microsoft/UserRepGen-4B
- https://huggingface.co/Qwen/Qwen3.5-4B
- https://localmodelwatch.tsuchitsuchi.com/2026/09/24/tev1-4b-experimental/
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: bartowski/UserRepGen-4B-GGUF, mradermacher/UserRepGen-4B-GGUF
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of bartowski/UserRepGen-4B-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

