# deepseek/deepseek-v3.2 across inference providers

Reference: context 163,840 tokens from https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/config.json (max_position_embeddings); quantization FP8 e4m3, 128x128 blocks.

Provider rows: https://openrouter.ai/api/v1/models/deepseek/deepseek-v3.2/endpoints, read 2026-10-19. Declared values; no requests sent.

| Provider | Tag | Quantization | Context | Max output | Tools | Reasoning | Differs from the reference |
|---|---|---|---|---|---|---|---|
| AtlasCloud | atlas-cloud/fp8 | fp8 (same) | 163,840 | 147,456 | yes | yes | none declared |
| Baidu | baidu/fp8 | fp8 (same) | 131,072 | 65,536 | no | yes | **declares a 131,072-token context; the reference config says 163,840 (80%)**; **declares no tools parameter; the official model supports tool calling** |
| DeepInfra | deepinfra/fp4 | fp4 (different) | 163,840 | 16,384 | yes | yes | **declares fp4 weights; the official release is FP8 e4m3, 128x128 blocks** |
| DigitalOcean | digitalocean | none (undisclosed) | 163,840 | 147,456 | no | yes | **declares no tools parameter; the official model supports tool calling** |
| Friendli | friendli | none (undisclosed) | 163,840 | 147,456 | yes | yes | none declared |
| GMICloud | gmicloud/fp8 | fp8 (same) | 163,840 | 147,456 | yes | yes | none declared |
| Google | google-vertex | none (undisclosed) | 163,840 | 65,536 | yes | yes | none declared |
| Mara | mara | none (undisclosed) | 32,768 | 7,168 | no | yes | **declares a 32,768-token context; the reference config says 163,840 (20%)**; **declares no tools parameter; the official model supports tool calling**; **declares at most 7,168 output tokens, which the reasoning tokens also count against** |
| Phala | phala | none (undisclosed) | 163,840 | 64,000 | yes | yes | none declared |
| SambaNova | sambanova | none (undisclosed) | 32,768 | 7,168 | no | yes | **declares a 32,768-token context; the reference config says 163,840 (20%)**; **declares no tools parameter; the official model supports tool calling**; **declares at most 7,168 output tokens, which the reasoning tokens also count against** |
| SiliconFlow | siliconflow/fp8 | fp8 (same) | 163,840 | 147,456 | yes | yes | none declared |
| Venice | venice | none (undisclosed) | 160,000 | 32,768 | yes | yes | declares a 160,000-token context; the reference config says 163,840 (98%) |

## Changes since tracking began

None yet.