Novita Labs has officially announced the training and release of DSpark speculative models designed specifically for Kimi-K2.6 and Kimi-K2.7-Code on the Hugging Face platform. This new solution comes with native serving support directly within the vLLM optimization library, promising to deliver efficient inference acceleration for the open-source community.
Detailed Developments
According to Novita Labs on July 13, 2026, the DSpark speculators are designed to address latency issues in the text generation process of Large Language Models (LLMs). The direct integration into vLLM allows developers and enterprises to easily deploy this solution into existing infrastructure without complex configurations. This is a strategic move to optimize hardware costs when running next-generation AI models from Kimi.
Technical & Technology Analysis
The speculative decoding technology applied in DSpark works by using a smaller, faster draft model to predict upcoming tokens, which are then validated in parallel by the target large model (such as Kimi-K2.6). Through real-world evaluation across six standard benchmarks with a batch size of 1, the system recorded impressive performance gains. Specifically, Kimi-K2.6 achieved a 2.55x average throughput increase (a 155% boost), while the code-focused variant Kimi-K2.7-Code reached a 2.36x average speedup.
Expert Opinions & Assessments
Technical experts note that the out-of-the-box integration into vLLM is a massive advantage for Novita Labs' solution. The vLLM library is currently one of the most popular LLM serving engines globally due to its smart memory management (PagedAttention). The combination of DSpark's optimized speculative algorithm and vLLM's resource management will significantly lower operational costs per token for commercial AI systems.
Impact & Future
For the AI developer community in Vietnam, the availability of open-source optimization tools like DSpark on Hugging Face lowers hardware resource barriers. Doubling inference speeds without degrading output quality opens up opportunities to deploy chatbots, virtual assistants, and code generation tools that run more smoothly on mid-range GPU setups, reducing reliance on expensive high-end server clusters.