Model reference · open weights
academic-ds is an open-weight language model from ByteDance. academic-ds-9B (BF16) weighs 18.7 GB; the smallest configuration that runs it is 2× RTX 3060 12 GB.
What it is
| Released by | ByteDance |
|---|---|
| Released | 2025-04-09 |
| Parameters | 9.4B |
| VRAM | 18.7 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory |
|---|---|---|---|
| 8K each | |||
| RTX 3060 12 GB … RTX 4060 Ti 16 GB 2 smaller cards | — | — | |
| RTX 3090 24 GB | 21 | all 8K | 23.4 GB |
| RTX 4090 24 GB | 21 | all 8K | 23.4 GB |
| RTX 5090 32 GB | 72 | all 8K | 31.0 GB |
| L40S 48 GB | 157 | all 8K | 44.0 GB |
| A100 80 GB | 384 | all 8K | 78.2 GB |
| H100 80 GB | 343 | all 8K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 447 | all 8K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 537 | all 8K | 107 GB |
| H200 141 GB | 740 | all 8K | 138 GB |
| B200 180 GB | 971 | all 8K | 176 GB |
| 2× RTX 3060 12 GB tensor parallel | 5 | all 8K | 11.6 GB a card |
| 2× RTX 4060 Ti 16 GB tensor parallel | 30 | all 8K | 15.4 GB a card |
| 2× RTX 4090 24 GB tensor parallel | 83 | all 8K | 23.4 GB a card |
| 2× RTX 3090 24 GB tensor parallel | 83 | all 8K | 23.4 GB a card |
| Requests at once | 8K tokens each |
|---|---|
| 1 | 20.3 GB |
| 5 | 20.9 GB |
| 8 | 21.3 GB |
| 16 | 22.5 GB |
| 32 | 25.0 GB |
| 64 | 29.8 GB |
One card, with vLLM's small-card settings.
From the model card
This is a 9B model whose architecture is deepseek-v3, trained from scratch using 350B+ tokens from fully open-source, English-only datasets. It is designed for development and debugging purposes within the open-source community.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.