DeepSeek V4 + Open Source Code is Insane: 1.6T Parameters, MIT License, and a 1/20th Price Tag

TamilTech AI summary

DeepSeek just dropped DeepSeek V4-Pro, a 1.6 trillion parameter Mixture-of-Experts model with a 1 million token context window, plus a lighter V4-Flash variant, both fully open-weight under an MIT license on Hugging Face with no signup walls. The big deal is that the API sits at roughly one-twentieth the price of Claude Opus while staying within striking distance on coding benchmarks like SWE-bench, and a new hybrid CSA plus HCA attention design slashes long-context compute and KV cache so serious workloads fit on far less hardware. For Indian startups, founders, and anyone watching API bills, this flips the old “open is for tinkering, closed is for production” story because you can self-host commercially, fine-tune for banking or regional languages, and keep data on Indian soil under RBI or DPDP rules. Flash is small enough to run quantized on a single H100, while Pro needs something like an 8x H200 node or a rental cluster from local providers if you are not ready for capex yet. If you want to try it today, point your OpenAI-compatible SDK at DeepSeek’s API, pull the weights from Hugging Face with vLLM, or A/B test through OpenRouter and run your own evals—the cost-quality math has clearly shifted.

• Is DeepSeek V4 really free to use commercially?
• Can I run DeepSeek V4-Pro on a single GPU in India?
• How does DeepSeek V4-Pro compare to Claude Opus 4.7 on coding?
• Where can I download the DeepSeek V4 weights?

AI-assisted summary, checked by the TamilTech editorial team.