AI engineering teams adopting frontier reasoning models have long operated under a frustrating compromise: pay opaque closed-lab API bills or settle for sparse open weights stripped of training lineage. IFM challenges that vendor-lock dynamic with K2 Horizon, a coordinated family of six models spanning from edge deployment to enterprise clusters: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B.
"The models and code are released under the Apache 2.0 license."
Rather than dumping unannotated checkpoints, IFM pairs Apache 2.0 weights with intermediate training stages, explicit data-construction recipes, architecture specifications, mixture proportions, training code, fine-grained run logs, and standardized evaluation pipelines. For enterprise architects, this level of transparency turns what used to be a black-box commodity into a predictable, auditable foundation for proprietary in-house fine-tuning.
Edge Performance and Sparse Routing Mechanisms
Across mathematical benchmarks, reasoning chains, code generation, and agentic workflows, the sub-10B models—0.9B, 3.7B, and 7B—establish compelling performance tiers for edge hardware. The 0.9B model targets compute-starved micro-devices like smartwatches and AR glasses, while the 3.7B and 7B configurations handle local processing on mobile platforms without cloud routing overhead.
At the data center and workstation tier, IFM combines the dense 32B model with an MoE-based 36B-A4B variant. The 36B-A4B model uses a Mixture-of-Value-Attention mechanism to activate only 4B parameters per forward pass, delivering reasoning benchmarks comparable to far heavier dense networks while containing GPU memory footprints. At the top of the stack, the 375B-A23B flagship delivers enterprise-grade reasoning on private infrastructure, providing a viable alternative to proprietary token pricing.
Structural Consistency Across Model Deployments
The entire fleet shares a unified training regime, common tokenization baseline (with a specialized reduced vocabulary for the 0.9B variant), standardized evaluation tooling, and native quantization support. This architectural consistency eliminates the friction of managing fragmented toolchains when moving workloads between edge devices and centralized clusters.
By publishing intermediate checkpoints alongside reproducible data recipes, IFM exposes the exact operational mechanics of how reasoning capabilities emerge across parameter scales. For engineering leaders evaluating total cost of ownership, deploying an auditable open-weight fleet on dedicated compute increasingly undercuts long-term dependence on closed proprietary endpoints.