Primary use stated or inferred cautiously from official documentation.
M
MICROSOFT / microsoft/UserRepGen-4B
UserRepGen-4B
Microsoft checkpoint based on Qwen3.5-4B that turns text conversations into descriptions of the task and observed behaviour. Weights were uploaded to the official repository on October 14, 2026. It generates representations for a separate simulator: it does not itself produce user turns or execute their task. Although it retains the base model’s multimodal architecture, this release supports text only and was evaluated in English.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | microsoft/UserRepGen-4B |
|---|---|
| Model type | Text and reasoning |
| Access model | Open weights |
| License | MIT; original Qwen assets under Apache 2.0 |
| Deployment | Self-hosting with Transformers or vLLM |
| Release | 14 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text |
| Exit | Text |
| Context window | 8,192 prompt tokens; 9,216-token training sequence |
| maximum output | Up to 1,024 tokens in documented use |
| Reasoning | Reasoning mode disabled in the official example |
| Published tools | |
| Structured Outings | JSON with task and user_behavior; validate output |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Transformers · vLLM |
Supervised assistant research and reconstruction of authorized conversations. Fine-tuning uses rank-64 LoRA and online supervised learning with hindsight feedback. Training used 3,200 conversations from seven datasets; the separate evaluation uses 700 conversations. The example requires validating task and user_behavior fields and limits formatted input to 8,192 tokens and generation to 1,024. Do not assume the base model’s extended context.
Representations are neither anonymous data nor verified descriptions of a person: they may retain sensitive information, infer unsupported traits and facilitate conversation linkage. Do not use them for surveillance, impersonation, profiling without consent or consequential decisions. The model card still lacks the paper’s bibliographic reference; available results are Microsoft’s claims, not independent validation.
03 / CAPABILITIES
What it can do
Orientation
Supervised assistant research and reconstruction of authorized conversations. Fine-tuning uses rank-64 LoRA and online supervised learning with hindsight feedback. Training used 3,200 conversations from seven datasets; the separate evaluation uses 700 conversations. The example requires validating task and user_behavior fields and limits formatted input to 8,192 tokens and generation to 1,024. Do not assume the base model’s extended context.
Context
8,192 prompt tokens; 9,216-token training sequence · Up to 1,024 tokens in documented use
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The model card declares an MIT licence for the model and retains Apache 2.0 for original Qwen assets. No commercial API or token rate has been verified. Open weights still require compute, memory and maintenance; the CPU smoke test is not a performance measurement. Pin a revision and review licences and permission to process conversations before deployment.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Microsoft reports simulation fidelity of 4.95 best-of-1 and 6.02 best-of-3 on a 1–10 scale. The protocol uses GPT-5.4 mini as simulator, assistant and judge, with three rollouts per representation. This is neither percentage accuracy nor a general-purpose benchmark. Inferama has not reproduced it; fidelity depends on simulator and data, and synthetic interactions do not replace evidence from real users.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| UserRepGen Simulation Fidelity (best-of-1) | 4,95 / 10 | Simulation fidelity | View source ↗ |
| UserRepGen Simulation Fidelity (best-of-3) | 6,02 / 10 | Simulation fidelity | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
UserRepGen-4B
Public weights, configuration, example and model card; paper reference pending.
07 / SOURCES