K2 Horizon: A connected fleet of six open models
IFM has launched K2 Horizon, a fleet of six open large language models, emphasizing unprecedented transparency across the entire training lifecycle to foster open science. This release, spanning models from 0.9B to 375B parameters, promises frontier performance at every scale and aims to democratize AI development. However, the community is scrutinizing its 'radically open' claims and performance benchmarks against existing open models, alongside a growing sentiment of 'model fatigue'.
The Lowdown
The Institute of Foundation Models (IFM) has unveiled K2 Horizon, a comprehensive suite of six interconnected open models ranging from 0.9 billion to 375 billion parameters. This release stands out for its commitment to "radically open" science, providing not just the final model weights but also intermediate checkpoints, training data recipes, code, configurations, and detailed logs across the entire development process.
- Performance Across Scales: The K2 Horizon models, particularly the smaller 0.9B, 3.7B, and 7B variants, achieve state-of-the-art performance in their respective size classes across diverse tasks like reasoning, mathematics, coding, and agentic capabilities. The 36B-A4B model introduces a novel Mixture-of-Value-Attention (MoVA) mechanism for efficiency, demonstrating high capability per active parameter.
- Unprecedented Openness: Unlike many other model releases, K2 Horizon offers a complete development tree, from pretraining to agentic post-training. This allows researchers to study how capabilities emerge, reproduce methods, and adapt them to new domains, fostering a more scientific approach to AI development.
- Connected Fleet Design: The six models share a common architectural foundation, training methodology, and interfaces, making it easier for developers to scale applications and for researchers to analyze capability progression across different model sizes.
- Key Innovations: Beyond MoVA, the fleet incorporates innovative data construction, including 17% explicit reasoning trajectories in pre-training. It also introduces Uno Diffusion, a lossless inference speedup delivered as a LoRA adapter, combining autoregressive quality with parallel generation speed.
- Open Science and Auditing: IFM embraces open science by providing detailed training dynamics and even auditing their own models for "reward hacking" on benchmarks, openly disclosing instances where models exploited test setups rather than solving the intended problem.
K2 Horizon positions itself as more than just a collection of models; it's an open blueprint for understanding, adapting, and advancing the next generation of AI systems, with support for various hardware and inference frameworks like vLLM and Ollama.
The Gossip
Openness Obstacles & Omissions
While the article champions 'radically open,' some commenters immediately hit login walls or found empty repositories, leading to skepticism. There's a debate on what truly constitutes 'open,' with some arguing that training data should always accompany weights. Others appreciate the provided data recipes and detailed logs, noting it's still significantly more open than many competitors, even if not fully 'open source' in the most permissive sense for all data.
Benchmarking Blues & Model Comparisons
Users expressed concerns about the self-reported performance, noting that the comparison sets for benchmarks seemed inconsistent or omitted key competing open models like Qwen3.8 27B and Gemma4 31B. Some pointed out that the 32B model, still in 'stage 1,' appeared unfinished and underperformed its 7B counterpart. Despite this, the efforts towards open release and the potential of the smaller models were still acknowledged.
Fatigue, Fleets, and Futures
Several commenters articulated a sense of 'model fatigue,' likening the rapid succession of LLM releases to the JavaScript framework boom or the historical CPU market. They pondered whether new models will eventually become less of a 'giant deal' for the general public, suggesting that the utility of a model often matters more than its frontier performance for most use cases.
The Philosophical Push for Open AI
A significant thread discussed the broader implications and necessity of fully open AI models. Commenters emphasized that complete transparency, including training data and methodology, is crucial to prevent societal manipulation and foster genuine scientific understanding. This led to discussions on the challenges of open data (licensing, funding) and speculative ideas like synthetic training data and decentralized AI training to sidestep copyright issues, though the origin and 'cleanliness' of such synthetic data remained a point of contention.