ds1-fukuroRetrieval that runs on CPU.
- Task
- Retrieval
- Hardware
- CPU. No GPU
- Batch size
- 200,000 tokens
- Texts per request
- 32
model="ds1-fukuro"terminal
Same vectors. A fraction of the wait.
DS1as low as 0ms
Cohere0ms
OpenAI0ms
p99 latency. 10× faster than OpenAI text-embedding-3-small.
10×
Two frontier models. Both fast.
ds1-miru-int8Code retrieval, squeezed to int8.
- Task
- Code retrieval
- Hardware
- CPU. No GPU
- Batch size
- 2,000,000 tokens
- Texts per request
- 2,048
model="ds1-miru-int8"Already on OpenAI embeddings? Change two lines.
Swap the base URL and the model, both highlighted below. Use your Takara key. The rest of your code stays as it is.
openai package
from openai import OpenAI
client = OpenAI(
api_key="your_api_key",
base_url="https://infer.takara.ai/v1",
)
response = client.embeddings.create(
model="ds1-fukuro",
input="your text here",
)
print(response.data[0].embedding)
print(response.model)応答速度
10×
faster embedding latency than OpenAI text-embedding-3-small.
コスト
70%
lower embedding costs than OpenAI text-embedding-3-small.
GPU不要
0
GPUs required to run DS1.
Don't trust our numbers? Time DS1 yourself.
0ms
From your browser. Median of 5 round trips.