IFM introduced the K2-Horizon-32B-FP8 language model

Company IFM released K2-Horizon-32B-FP8 — a dense language model with 32 billion parameters. The release is focused on programming tasks and logical reasoning, demonstrating improved performance compared to existing analogs in its class.

The model's architecture supports a context window of 512 thousand tokens. To optimize computational resources and reduce memory requirements, developers applied quantization in the FP8 format. This allows the model to be run on less powerful equipment without significant loss of calculation accuracy.

According to internal tests, K2-Horizon-32B-FP8 surpasses Qwen3.8-27B in coding tasks and complex reasoning. The model shows stable performance when processing long sequences, making it suitable for analyzing large repositories and technical documentation.

The model's technical specifications and weights are already available for download on the Hugging Face platform. The use of the FP8 format ensures a balance between inference speed and generation quality, which is critical for deploying high-load systems.

The release of this model expands IFM's range of solutions, offering an alternative for developers who require high accuracy when working with code. The transition to quantized formats is becoming a standard for the efficient use of 32-billion models in production environments.