IFM releases K2 Horizon, a six-model open fleet from 375B down to 0.9B

IFM has released K2 Horizon, a connected family of six language models stretching from a 375-billion-parameter flagship down to a 0.9-billion-parameter model aimed at wearables. According to the company's announcement, which surfaced on the front page of Hacker News, the three smallest sizes claim new state-of-the-art results for their classes, and every model in the fleet ships with its full training pipeline attached.

One fleet from edge to enterprise

K2 Horizon consists of the 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B variants. IFM maps them onto the full deployment spectrum: the 0.9B model targets hardware as constrained as smartwatches and glasses, the 3.7B and 7B models aim at phones and other on-device scenarios, the dense 32B and sparse 36B-A4B models fit local workstations and efficient serving setups, and the 375B-A23B handles large enterprise deployments. All six ship with quantization support.

The suffixes describe the architecture. The 375B-A23B is a sparse mixture-of-experts model that stores 375 billion parameters but activates roughly 23 billion per token, while the 36B-A4B activates about 4 billion. That mid-size model also debuts what IFM calls Mixture-of-Value-Attention (MoVA), a sparse attention design the company credits with delivering performance close to the dense 32B model at a fraction of the active compute. IFM further frames the 32B and 36B-A4B pair as a controlled reference point for comparing dense and sparse architectures trained under similar conditions.

Because the six models share core architecture, vocabulary (reduced for the 0.9B), training methodology, interfaces, evaluation infrastructure and deployment tooling, IFM argues the fleet makes it simpler to shift workloads between sizes and to study how capability and efficiency change with scale.

The benchmark picture

Every performance figure comes from IFM itself. The company says the 0.9B, 3.7B and 7B models lead their size classes across mathematics, reasoning, general capability, coding and agentic evaluations, with the 0.9B scoring above 48 on AIME 2026. The 3.7B and 7B models are reported to perform well on SWE-bench and BrowseComp, and on several evaluations they land near or above figures from previous-generation models several times their size. The 32B ranks among the top dense models below 40 billion parameters, and the 375B-A23B among the strongest models below 400 billion overall.

The company is upfront about the limits: benchmarks that require long chains of trial, error and recovery, such as TerminalBench, remain difficult for the smallest models in the fleet.

What is actually being opened up

IFM describes K2 Horizon as its most comprehensive open release to date and claims it is the first open model family to carry transparency all the way through agentic post-training. For each of the six models, the bundle covers intermediate checkpoints from pretraining onward, the training code and configurations, mixture recipes, detailed logs, evaluation results and final weights, together with the training data itself or, where the data cannot be redistributed, a thorough account of how it was constructed and mixed. Models and code are licensed under Apache 2.0, while datasets fall under their applicable licenses such as ODC-BY.

The flagship is published as what IFM calls a development tree rather than a single endpoint: checkpoints and post-training branches are intended to show how the base model evolves into reasoning, instruction-following and specialized agentic variants.

Why it matters

The release targets a familiar trade-off. Fully documented open models tend to sit well behind the capability frontier, while frontier-grade models released as weights alone reveal nothing about how their abilities were built. By pairing competitive models with the recipes that produced them, IFM gives researchers systems strong enough to actually display reasoning, tool use and planning to study, and gives developers a path to reproduce and adapt those methods to new tools and domains instead of starting from an opaque checkpoint.

The small end of the fleet deserves attention as well. If the claims survive independent evaluation, sub-10B models with credible agentic behavior would push multi-step browsing and software-engineering tasks onto phones and wearables rather than cloud GPUs. IFM traces its fully open approach to its 2023 LLM360 paper and says it has released open models every year since, with K2 Horizon extending that commitment across the complete training lifecycle.