DeepSeek's DSpark: Making LLMs Faster Without Touching the Model"

 the tool, its open-source MIT release, the speculative decoding mechanism, and the general performance direction are all confirmed.

 presenting "85% faster" as a single clean number rather than a range with caveats; the analogy is a fair simplification but omits that these are unverified vendor benchmarks; and speculative decoding itself is framed as if DeepSeek invented the "smarter way to run LLMs" concept, when it's an existing technique DeepSeek has optimized and productionized.