Model comparison
UI-TARS-72B-DPO vs gemma-4-26B-A4B-it-qat-q4_0-unquantized
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: gemma-4-26B-A4B-it-qat-q4_0-unquantized's KV cache at 32K is 6.5× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| UI-TARS-72B-DPO | gemma-4-26B-A4B-it-qat-q4_0-unquantized | |
|---|---|---|
| Parameters | 73.4B | 26.5B |
| Architecture | qwen2vl | gemma4 |
| Layers | 80 | 30 |
| Native context | 32,768 | 262,144 |
| Mixture of experts | no | yes, 128 experts |
| Quantizations published | 26 | 1 |
| Smallest quantization | 22.11 GiB | 13.45 GiB |
| Q4_K_M | 44.16 GiB | — |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | UI-TARS-72B-DPO | gemma-4-26B-A4B-it-qat-q4_0-unquantized | Ratio |
|---|---|---|---|
| 4,096 | 1.25 GiB | 0.45 GiB | 2.78× |
| 8,192 | 2.50 GiB | 0.61 GiB | 4.13× |
| 16,384 | 5.00 GiB | 0.92 GiB | 5.45× |
| 32,768 | 10.00 GiB | 1.54 GiB | 6.48× |
| 65,536 | 20.00 GiB | 2.79 GiB | 7.16× |
| 131,072 | 40.00 GiB | 5.29 GiB | 7.56× |