According to a recent LinkedIn post from FriendliAI, the company is highlighting support for draft-model speculative decoding in its Dedicated Endpoints for large language model inference. The post describes a setup where draft models are trained and automatically paired by FriendliAI and can be enabled with a single configuration toggle, without requiring application code changes.
Meet Samuel – Your Personal Investing Prophet
- Start a conversation with TipRanks’ trusted, data-backed investment intelligence
- Ask Samuel about stocks, your portfolio, or the market and get instant, personalized insights in seconds
The LinkedIn post suggests this approach is designed to accelerate generation by predicting multiple tokens ahead and verifying them in parallel, aiming to reduce the autoregressive bottleneck that typically slows longer outputs. It also notes that verification follows the target model’s own next-token distribution, which the company implies can preserve output quality while maintaining similar computational cost per step.
According to the post, this speculative decoding method may be particularly suited to agentic pipelines, long-form text generation, and code completion, all of which are workloads of growing commercial interest in AI infrastructure. The post also contrasts draft-model speculative decoding with N-gram-based methods, indicating that the former could generalize beyond literal token repetition and potentially offer broader performance gains.
The post lists supported target models as including Gemma-4-31b-it, Kimi-K2.6, Qwen3.6-27B, GLM-5.1, GLM-5, DeepSeek-V3.2, and MiniMax-M2.5, suggesting FriendliAI is targeting users of several popular open and regional foundation models. For investors, this focus on latency reduction and ease of integration could strengthen FriendliAI’s positioning in the AI infrastructure market, as faster, higher-throughput inference is a key driver of cloud cost efficiency and customer adoption.
If the described performance and integration benefits prove material at scale, FriendliAI may be able to deepen relationships with AI-native enterprises and developers seeking to optimize inference workloads without major engineering changes. This could support higher usage-based revenue, improve competitive differentiation against other inference platforms, and enhance the company’s strategic relevance as generative AI applications continue to expand.