DeepSeek Releases V4 DSpark Models With Faster Speculative Decoding
DeepSeek released DeepSeek-V4-Pro-DSpark, with 1.6T total and 49B active parameters, and V4-Flash-DSpark on Hugging Face under the MIT license. Compared with MTP-1, DSpark improved per-user generation speeds by 57–78% for Pro and 60–85% for Flash. Both support a 1M-token context. The framework was announced June 27, 2026, and the models ranked highly on Hugging Face Trending on July 3–4.
Visit original source