ModelsModels / DeepSeek
DeepSeek-V4-Pro
Gets 94 of every 100 tool calls right on the public test. Best at calling a single tool, weaker at choosing the right tool from several.
DeepSeek-V4-Pro is made by DeepSeek. It takes text in and answers in text, with a context window of 1M tokens.
Start using DeepSeek-V4-Pro in 3 steps
Read the docs- Get an API keyFree to create. It works for every model in the catalogue.Get an API key
- Add creditsPay as you go. Credits work on every model.Add credits
- Point your client at TīrthaAny OpenAI-compatible client works. Use your key and this base URL, and name the model.base_url: "https://api.tirtha.ai/v1" api_key: "YOUR_TIRTHA_KEY" model: "catalog/deepseek-v4-pro"
Price
The live price, read from billing as this page loads. You pay only for the tokens you use, from your credits.
At a glance
Tested on Tīrtha Bench
Tool calling (public BFCL v4 test)
Overall: 94.1%*. Picks the right tool and fills in the inputs correctly on 94 of every 100 tasks.
One tool 96% · Picking the right tool 92% · Several calls at once 92%
A task counts as right only if the model calls the right tool with the right inputs. 800 tasks. Run on Oct 29, 2026. *Measured by Spanda Works on third-party hosts, not through Tīrtha. Results through Tīrtha may differ.
Handled several calls at once in our tests.
See DeepSeek-V4-Pro beside every other model on Tīrtha Bench →
What it supports
- Tool callingYesAlso measured on Tīrtha Bench
- Text inYes
- ReasoningYes
- StreamingYes
- Sampling settingsFixed todayThe maker suggests temperature 1, top p 1. Tīrtha fixes temperature at 0 today.
From the maker
- Model card: deepseek-ai/DeepSeek-V4-Pro on Hugging Face
- Last updated by the maker: Jun 22, 2026
Facts here are read from the maker's published card; claims and benchmarks on that card are the maker's own, not measured by Tīrtha.
Licence
Weights: MIT. Read the licence.