Here's what breaks first at a 1M-token context: the thing you invented to survive it.
Sparse attention needs an indexer to pick which tokens matter, and if all 78 layers run their own indexer over a million tokens, the index eats the win. Hy4-preview's answer is IndexCache.
A blind writing test produced the result Mark least expected
DeepSeek and Qwen beat the frontier models he normally reaches for. Cheap did not just compete. It won his preference test. That is enough to rethink your default model choices.
Local AI is amazing. One DGX spark is cool, 2 DGX sparks even better!
This first video is Deepseek using MiaAI_lab recipe on 1 spark. The second video is using the Glm recipe for 2 sparks. Both given the same prompt (in comments). Both 1-shot.
Deepseek harness tui makima-tui
Git：
You have reached the end of the archive
All of deepseek