IFM Releases K2-Horizon-7B with Plug-and-Play Diffusion Adapter for 5,200 TPS Inference
IFM has released K2-Horizon-7B, a 7-billion parameter causal language model augmented with a plug-and-play diffusion adapter alongside its autoregressive weights. The model achieves inference speeds of up to 5,200 tokens per second while claiming zero output quality loss. Achieving up to 5,200 tokens per second without quality degradation represents a major step forward for real-time AI applications and non-autoregressive decoding techniques. If broadly validated, combining causal architectures with diffusion adapters could drastically cut inference latency and operational costs for high-throughput systems. The model attaches a modular diffusion adapter to standard causal transformer weights, allowing parallel or accelerated token generation rather than strictly sequential decoding. The model artifacts are available on Hugging Face alongside technical details provided in paper arXiv:2609.04010.
## BACKGROUND
Standard large language models generate text autoregressively by predicting one token at a time, which creates a latency bottleneck during inference. In contrast, diffusion models generate outputs by iteratively refining noise into structured data, allowing broader parallel generation. Plug-and-play adapters enable developers to add modular capabilities to existing base models without retraining the core parameters from scratch.