FreeLLMAll models › microsoft
UserRepGen-4B
microsoft · released 14 Oct 2026
Download New · waiting for Arena votes
Arena score—Not voted on Arena yet
Size4.5 Bparameters
4-bit download2.7 GBweights only, Q4_K_M
Max context256Ktokens, from the model config
Where to use it for free
- Download and run it at home: official weights · GGUF files for Ollama and LM Studio
New: waiting for Arena votes
Released on 14 Oct 2026. Arena has not voted on this model yet. As soon as it does, the model moves into the ranking automatically, usually the next night. Until then you can still check its size, memory needs and speed below.
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 69 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 56 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 161 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 171 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 269 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 165 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 156 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 148 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 80 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 50 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 337 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 78% ÷ data read per token, plus a fixed cost per token, calibrated on 44 real llama.cpp tests; typical error about ±10%. Compare all GPUs.
GPU memory needed
| Level | 8K context | 32K context | 128K context | Quality vs full |
|---|---|---|---|---|
| Q3_K_M | 3.1 GB | 3.9 GB | 7.1 GB | +10.5% noticeable loss |
| Q4_K_M | 3.6 GB | 4.4 GB | 7.6 GB | +2.8% small loss |
| Q5_K_M | 4.1 GB | 4.9 GB | 8.1 GB | +0.9% very small loss |
| Q6_K | 4.6 GB | 5.4 GB | 8.6 GB | +0.3% near-lossless |
| Q8_0 | 5.6 GB | 6.5 GB | 9.7 GB | ≈ 0% same as full |
| F16 | 9.9 GB | 11 GB | 14 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.