Model comparison
HunyuanOCR vs embeddinggemma-300m
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: embeddinggemma-300m's KV cache at 32K is 5.8× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| HunyuanOCR | embeddinggemma-300m | |
|---|---|---|
| Parameters | 1.1B | 303M |
| Architecture | hunyuan_vl | gemma-embedding |
| Layers | 24 | 24 |
| Native context | 131,072 | 2,048 |
| Mixture of experts | yes, 1 experts | no |
| Quantizations published | 25 | 10 |
| Smallest quantization | 0.24 GiB | 0.26 GiB |
| Q4_K_M | 0.33 GiB | — |
| Licence | — | — |
KV cache by context
the term that decides long-context viability
| Context | HunyuanOCR | embeddinggemma-300m | Ratio |
|---|---|---|---|
| 4,096 | 0.11 GiB | 0.04 GiB | 3.00× |
| 8,192 | 0.21 GiB | 0.05 GiB | 4.15× |
| 16,384 | 0.42 GiB | 0.08 GiB | 5.14× |
| 32,768 | 0.84 GiB | 0.14 GiB | 5.84× |
| 65,536 | 1.69 GiB | 0.27 GiB | 6.26× |
| 131,072 | 3.38 GiB | 0.52 GiB | 6.50× |