Skip to content
X · @Yuchenj_UW

Finally, DeepSeek V4 is here!

DeepSeek announced DeepSeek V4 and says DeepSeek-V4-Pro uses an MIT license with 1.6T parameters and 49B active parameters. The snippet also claims DeepSeek-V4-Pro Max is close to Opus-4.6 Max and GPT-5.4 xHigh across benchmarks; the post does not disclose benchmark names, scores, release timing, or model weights. The key signal is the MIT license and 49B active scale, not the headline comparison.

Why it matters: DeepSeek open-sourced V4 Pro under MIT with 1.6 trillion total parameters and only 49 billion active at inference, setting a new reference for open-model deployment cost.

Read the original ↗Export Markdown