IFM Releases K2-Horizon-375B-A23B-NVFP4 Quantized Model

September 23, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

IFM has released IFM/K2-Horizon-375B-A23B-NVFP4, an NVFP4 quantized version of the sparse Mixture-of-Experts (MoE) model IFM/K2-Horizon-375B-A23B. This model aims to reduce memory usage and speed up inference in environments that natively support NVFP4 on NVIDIA Blackwell generation (B-series) GPUs and newer.

Quantization is applied exclusively to the weights and activations of the routed-expert linear layers, while all other linear layers—such as attention, shared experts, routers, the first three dense layers, and lm_head—remain in BF16.

Specifications

• Total parameters: 375B
• Architecture: MoE (Total parameters 375B, Activated parameters 23B)
• Context length: 512K

Performance

According to measurements by the publishers, the NVFP4 version shows a slight performance decrease compared to the original BF16 version, but reduces memory usage and enables faster inference on NVFP4-supported hardware. Note that evaluation of the NVFP4 version has currently been conducted only on non-agent tasks, with agent task results scheduled to be released at a later date.

Comparison between the NVFP4 version and the BF16 version (columns narrowed):

The benchmark results of the base model K2-Horizon-375B-A23B compared with other models are shown in the table below (columns narrowed):

According to measurements by the publishers, this model demonstrates agent performance comparable to or exceeding open-weight MoE models 2.6 times its size, and is competitive with closed frontier models. Specifically, it records high scores in agentic tasks such as MCPMark (67.7) and Toolathlon (65.3). On the other hand, it falls short of MiniMax-M3 (92.9) in GPQA Diamond (87.3), and a difference is also observed when compared to MiniMax-M3 (83.5) and Claude Sonnet5 (84.7) in BrowseComp (72.8). Additionally, it shows lower scores than the comparison models in physics reasoning tasks like CritPt (8.6).

Strengths and Use Cases

The base model K2-Horizon-375B-A23B is characterized by its high performance as an agent. According to the model card, it exhibits performance comparable to or exceeding open-weight MoE models 2.6 times its size in benchmarks for agentic tool use, terminal operation, and long-term workflows, remaining competitive with closed frontier models. It demonstrates high capability particularly in agentic tasks such as MCPMark (67.7) and Toolathlon (65.3), as well as terminal operation in Terminal-Bench 2.1 (70.2).

It also natively supports a very long context window of 512K, making it suitable for tasks requiring long-document reading comprehension and complex context. In the coding domain, it shows solid performance in repository-level tasks such as SWE-Atlas-QnA (48.4) and SWE Bench Pro (42.6).

This NVFP4 quantized version specializes in maintaining these high capabilities while achieving reduced memory usage and faster inference on NVIDIA Blackwell generation hardware and newer.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 218.2B parameters

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-09-22): llama.cpp: not registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Recent Models in the Same Size Class

Models with over 40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

How to Get It

• Distribution format: safetensors (NVFP4 quantized version)
• Download command: huggingface-cli download IFM/K2-Horizon-375B-A23B-NVFP4
• Supported engines: vLLM (recommended recipe available), SGLang (recommended recipe available)

This model is released under the Apache 2.0 license.

Related Articles

• NVIDIA Releases NVFP4 Quantized DeepSeek-V4-Pro
• OpenBMB Releases MiniCPM5-2B: A SOTA 2B On-Device Model

Sources

• https://huggingface.co/IFM/K2-Horizon-375B-A23B-NVFP4