Institute of Foundation Models launches K2 Horizon 375B fleet

On September 3, 2026, the Institute of Foundation Models (IFM) unveiled K2 Horizon, a groundbreaking fleet of six AI foundation models ranging from 0.9 billion to 375 billion parameters. This release stands as the largest fully open-source model launch in AI history, providing researchers and developers with complete access to weights, code, training data, and methodologies. The models are optimized for diverse deployment scenarios, from edge devices like smartwatches to enterprise-scale servers, with the 375B-A23B variant targeting high-demand enterprise applications.

While OpenAI’s GPT-6 Astra dominated headlines this week, CNET reported that Anthropic, Meta, and Google also rolled out updates, emphasizing advancements in agentic AI workflows. Unlike these competitors, which prioritize performance and proprietary features, IFM’s K2 Horizon prioritizes transparency, enabling full reproducibility and adaptation,a stark contrast to the industry’s ongoing debate over “open weights” versus comprehensive openness.

What sets this release apart is its radical transparency: every model includes training data, recipes, and evaluation metrics, allowing researchers to inspect and replicate results without barriers. This approach challenges the trend of closed ecosystems and signals a shift toward democratizing AI development. By contrast, most recent launches, including those from OpenAI and Google, focus narrowly on capability rather than openness, leaving critical questions about reproducibility and ethical training practices unresolved.

The article will explore how IFM’s commitment to full openness could reshape collaboration in AI research, contrasting it with the proprietary strategies of industry giants. Read more about the K2 Horizon launch here.

On September 3, 2026, the Institute of Foundation Models unveiled the K2 Horizon fleet, a collection of six fully open‑source AI models ranging from 0.9 billion to 375 billion parameters finance.yahoo.com. The announcement emphasized that all models, including weights, code, training data and methodology, are publicly released, marking the largest open‑source model launch in AI history. Researchers can inspect, reproduce and adapt each model without licensing restrictions. This transparency is expected to accelerate scientific progress and democratize access to state of the art capabilities.

The six models share a unified core architecture, common vocabulary (except the 0.9 B variant), identical training methodology and standardized deployment tooling, as reported by CNET cnet.com. Dynamic routing directs inference tasks to the most economical model, enabling a seamless transition from prototype development on a small device to large‑scale production on enterprise hardware. This architecture allows the 0.9 B model, optimized for wearables, to coexist with the 375 B‑A23B behemoth designed for data‑center workloads. The unified stack simplifies integration for developers who can switch between models without rewriting code.

Prior to this release, most high‑performance models were offered only through closed APIs, limiting reproducibility and stifling academic research. By providing full openness across all size classes, IFM challenges the prevailing “open‑weights” narrative and sets a new benchmark for transparency in the industry. The move could accelerate collaborative benchmarking and reduce barriers for smaller organizations seeking advanced AI capabilities.

The K2 Horizon lineup includes a 0.9 B model tailored for ultra‑constrained devices such as smart watches and glasses, while the 3.7 B and 7 B variants are optimized for on‑device applications on smartphones and edge processors finance.yahoo.com. The dense 32 B model and the sparse 36 B‑A4B model target local servers and on‑premise deployments, delivering high performance for enterprise‑grade workloads. The flagship 375 B‑A23B model is engineered for demanding corporate environments that require massive reasoning, mathematical and coding capabilities. Each size class sets new state of the art benchmarks across reasoning, mathematics, coding and agentic tasks.

According to pricing data from pricepertoken.com, the 0.9 B model incurs roughly $0.06 per token, making it the most cost effective option for edge devices, whereas the 375 B model’s token cost rises to about $0.20, reflecting its higher compute demands pricepertoken.com. The dynamic routing system described by IFM automatically selects the optimal model based on workload and budget, enabling developers to move from low cost prototyping on a 0.9 B device to high performance inference on the 375 B server without manual intervention. This automated selection also reduces operational expenses for enterprises that would otherwise need to provision dedicated hardware for each model size. Early adopters report up to 40 % savings in compute costs when leveraging the fleet’s tiered architecture.

The breadth of model sizes in the K2 Horizon fleet illustrates a strategic shift toward modular AI deployment, where organizations can match computational resources precisely to task requirements. This flexibility contrasts with earlier monolithic releases that forced a single model size for all use cases. As the industry moves toward more sustainable and economical AI practices, the open‑source nature of the fleet may become a decisive factor for widespread adoption across sectors.

Democratizing Research Through Complete Reproducibility

On September 3, 2026, the Institute of Foundation Models unveiled K2 Horizon, a suite of six fully open‑source foundation models that span from 0.9 billion to 375 billion parameters finance.yahoo.com. Each model ships with its training data, recipes, and evaluation benchmarks, moving beyond the limited “open weights” approach that has dominated recent AI releases. “Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it,” said Eric Xing, IFM’s founder and president. This comprehensive transparency enables developers worldwide to inspect, reproduce, and adapt the models for their own workloads, fostering a richer ecosystem of capabilities and scales. The initiative positions K2 Horizon as a new benchmark for openness in an industry where full disclosure is still evolving.

Concurrent industry moves highlight a growing emphasis on accessibility and safety, with Anthropic’s cost‑reduced Fable 5.1 and OpenAI’s GPT‑6 Astra illustrating the trend toward more affordable and secure models cnet.com. These releases underscore a broader shift away from proprietary black‑boxes toward models that can be examined and refined by the community. K2 Horizon’s fully open fleet stands out by offering not just model weights but also the underlying data and training procedures, setting a higher bar for scientific reproducibility. As enterprises and researchers alike seek greater control over AI systems, the availability of complete toolkits becomes a decisive differentiator. The combined effect of cheaper alternatives and fully transparent models is likely to accelerate adoption across diverse technical environments.

The push for complete reproducibility democratizes AI research by removing barriers to entry for smaller teams and academic labs. When training data and recipes are openly shared, novel architectures can be validated more quickly, reducing duplication of effort and fostering collaborative innovation. This openness also encourages the emergence of niche models tailored to specific domains, which might otherwise be eclipsed by large, closed‑source systems. Over time, the cultural shift toward full transparency could reshape funding priorities, with more emphasis placed on open‑science practices rather than solely on proprietary advantage. As a result, the AI community may see a more equitable distribution of capabilities and a faster pace of breakthrough discoveries.

Why K2 Horizon matters now
The Institute of Foundation Models unveiled K2 Horizon, a fleet of six AI models ranging from 0.9 billion to 375 billion parameters. All models ship with full weights, source code, training data and methodology, making the release the largest fully open‑source model collection in history. The 0.9 billion variant targets edge devices such as smart glasses, while the 375 billion‑A23B model aims at enterprise‑grade workloads. By sharing the complete stack, IFM enables researchers to reproduce results and adapt models without licensing barriers IFM.

Developers now have a scalable open alternative to closed‑source offerings like GPT‑6 Astra, which dominates benchmarks but remains proprietary. The release comes amid a week of announcements from Anthropic, Meta and Google, yet none provide models of comparable size with open training data. This gap highlights a persistent industry trend where only a few firms control the largest models, limiting reproducibility and community driven innovation. The open nature of K2 Horizon accelerates research cycles and lowers entry barriers for startups and academic labs CNET.

Evertune’s AI model tracker shows 124 releases in August 2026, indicating a rapid pace of innovation that outstrips evaluation of long‑term impact. K2 Horizon’s size and openness stand out because most recent launches focus on modest parameter counts or incremental improvements rather than breakthrough scale. The sustainability of such a massive open fleet depends on funding and compute resources, and the industry must address these practical concerns. Nonetheless, the launch signals a decisive shift toward transparent, reproducible AI research at an unprecedented scale Evertune.

The Institute of Foundation Models unveiled K2 Horizon, a fleet of six fully open‑source foundation models ranging from 0.9 B to 375 B parameters. Each model ships with its weights, code, training data and detailed methodology, enabling full reproducibility. Performance benchmarks show the smallest variants set new state‑of‑the‑art for on‑device use while the larger versions excel in reasoning, coding and agentic tasks. A shared architecture and dynamic routing layer let developers pick the most cost‑effective model for any workload.

By releasing the largest open‑source model suite to date, IFM raises the bar for transparency and challenges competitors to move beyond open weights toward full data and code disclosure. The approach lowers barriers for researchers, startups and enterprises to experiment, fine‑tune and deploy models from edge devices to data centers. As more teams adopt the shared tooling, we may see faster innovation cycles and a shift in how model licensing is negotiated. Will this openness spark a new wave of collaborative AI development that redefines industry standards?

Frequently Asked Questions

What models are included in the K2 Horizon fleet?
The fleet consists of six models with 0.9 B, 3.7 B, 7 B, 32 B dense, 36 B‑A4B sparse and 375 B‑A23B parameters.

How does K2 Horizon ensure reproducibility?
Every release provides the model weights, training code, full dataset and detailed training recipes so anyone can replicate the results.

What performance advantages do the smaller K2 Horizon models offer?
The 0.9 B, 3.7 B and 7 B versions achieve top‑tier scores on reasoning, math and coding benchmarks for their size, making them suitable for watches, phones and other constrained devices.

Can enterprises deploy the largest K2 Horizon model on‑premise?
Yes, the 375 B‑A23B model is designed for demanding enterprise workloads and can be run on local servers or cloud instances using the provided deployment tooling.

How does the dynamic routing feature work?
The system automatically directs incoming tasks to the smallest model that meets the required accuracy, reducing cost while maintaining performance.