The release is for a three- or four-node DGX Spark setup and says serving kernels are now TileLang except for NCCL and a few rare-path ATen copies.
On a four-node ring, the notes report prefill gains of 11.8% to 13.9% versus spark-ds41f-1022 across 32K to 1M prompts. On a three-node triangle, they report 22.5% to 28.2% faster prefill, with 3,952 tokens per second at a 1M-token prompt. The four-node configuration reports a 16,384-token prefill budget and a vLLM startup estimate of 6,512,700 KV-cache tokens.
The release fixes a prompt_logprobs request that could bring down the engine and a speculative-decoding logprob-token-ID HTTP 500. Its test record lists 5,730 passed and 160 skipped tests, with none failed. The notes do not establish independent benchmark validation, performance on other hardware, or general availability as a supported DeepSeek serving product.