DeepGEMM: The Free Engine DeepSeek Runs Its AI On, and Why It Makes AI Cheaper
DeepSeek open sourced the low-level GPU math library it trains and runs its own models on. It's why the same computers can answer far more questions per second, and a big part of why DeepSeek's prices are so low. What it is, who can actually use it, and what it means for your AI bill
Every answer an AI model gives is millions of multiplications on a graphics card. Do that math faster on the same hardware, and every answer costs less to serve.
That's what DeepGEMM does. It's DeepSeek's own kernel library, the exact code it trains and runs its models on, and it's free and open source on GitHub under the MIT license.
What DeepGEMM is
- A library of the core math every large language model runs on, written for NVIDIA's newest data-center GPUs.
- Fast: DeepSeek reports up to 1,550 TFLOPS on a single H800, matching or beating libraries that experts tune by hand.
- Kept up to date: DeepSeek keeps adding the kernels behind its newest models, and on Sept 30 it released a version for Huawei's Ascend chips too.
- Free: MIT license, so anyone can use it, change it and ship it. Repo: github.com/deepseek-ai/DeepGEMM
Why faster math means cheaper AI
- A GPU costs the same per hour whether it answers ten questions or a thousand.
- Faster kernels mean more answers per second from the same GPU.
- More answers per GPU-hour means each answer costs less to serve.
- Cheaper serving is a big part of how a lab can charge a fraction of what others do.
You can see the gap in the prices, per million output tokens:
| Model | Output price per million tokens |
|---|---|
| DeepSeek V4-Pro | $3.96 at peak, $1.98 off-peak |
| Claude Haiku 4.5 | $5 |
| Claude Sonnet 5.5 | $10 |
| Claude Opus 5.5 | $20 |
| Claude Fable 5.1 | $50 |
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (2 more sections, 2 copy-paste prompts) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: Can you use it yourself? · No GPUs? Here's what it means for you
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.