Glonce

IFM releases K2 Horizon, six open models from 0.9B to 375B parameters

modelsThe Glonce editorsHacker News🔥78Glonce rating78Impact72Novelty68Relevance98Surprise48Players38Credibility87
IFM releases K2 Horizon, six open models from 0.9B to 375B parameters

IFM has released K2 Horizon, a connected fleet of six open models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B. The fleet spans deployment from constrained edge devices such as watches and glasses up to demanding enterprise workloads, and IFM says every size class delivers top-tier performance across reasoning, mathematics, coding, agentic tasks and general capability, with the 0.9B, 3.7B and 7B models setting new state-of-the-art results at their respective scales.

IFM calls this its most comprehensive open release to date, and the claim rests on more than final weights. For every model in the fleet, IFM is opening the entire training lifecycle, from pretraining through reasoning and agentic post-training: intermediate checkpoints, training data or detailed data-construction recipes, the open architecture, mixture compositions, training code, configurations, fine-grained logs and evaluation results. Models and code are released under the Apache 2.0 license; datasets are released under their own applicable licenses such as ODC-BY, and where a dataset cannot legally be redistributed, IFM discloses how it was constructed and mixed instead. IFM frames the release around a tradeoff it states directly: a transparent model that falls far behind the capability frontier has limited value as a foundation, even for research, while a powerful model released only as final weights lets people run it but gives little insight into how its capabilities were created.

K2 Horizon 375B-A23B is the fleet's largest and most capable model, a sparse mixture-of-experts design with 375 billion total parameters that activates about 23 billion of them for each token. IFM says it ranks among the top models below 400 billion parameters on general, reasoning, coding and agentic evaluations, and positions it for demanding workloads such as complex reasoning, software engineering, research and long-horizon agentic tasks. Like the rest of the fleet, it is released as a development tree rather than a single endpoint, with intermediate checkpoints and post-training branches exposed to show how the base model develops into reasoning, instruction-following and specialized agentic variants.

Horizon 32B is the fleet's most powerful dense model, ranking among the top dense models below 40 billion parameters and aimed at local deployment. Its sparse counterpart, Horizon 36B-A4B, holds 36 billion total parameters but activates only about 4 billion of them per token, using a new architecture called MoVA, short for Mixture-of-Value Attention. Under matched training conditions, IFM says 36B-A4B performs only slightly below the dense 32B model despite the much smaller active-parameter count, and it positions the pair as a reference point for studying how dense and sparse architectures behave under identical conditions.

MoVA extends the logic of mixture-of-experts, normally applied only to a model's feed-forward layers, into the attention mechanism itself: a router activates only a subset of experts for each token's attention computation instead of running every expert on every token. IFM says MoVA integrates expert routing into multi-head attention while staying compatible with established efficiency techniques including FlashAttention, grouped-query attention and sparse attention.

The three smallest models target on-device and edge use. K2 Horizon 7B and 3.7B are built for phones and similar on-device applications; IFM cites strong performance on SWE-bench and BrowseComp without publishing scores for either, and says that on several evaluations the two models' results approach or exceed those of previous-generation models many times their size. K2 Horizon 0.9B is designed for highly constrained environments such as watches and glasses under quantization; IFM reports an AIME 2026 score above 48 for it, alongside the ability to perform mathematical reasoning, use tools and complete simple agentic tasks. IFM notes that complex tasks requiring extensive exploration and repeated recovery, of the kind TerminalBench is built to test, remain difficult for the smallest models even as the fleet overall pushes the performance boundary at every scale.

Each model in the fleet was pretrained on about 20 trillion tokens, drawn from web, code, mathematical, scientific, multilingual and domain-specific sources mixed with synthetic data from IFM's own pipelines; across pretraining overall, IFM used roughly 10 trillion synthetic tokens. Close to 17% of the pretraining corpus consists of problem-solving trajectories with explicit reasoning, with math-focused trajectories further rewritten into formats such as dialogues and study guides. To measure how diverse that synthetic data really was at corpus scale, without hitting the saturation that conventional gzip or zstd compression metrics show as document counts grow, IFM built a new compressor called Wzip, combining an adaptive multi-sliding-window LZ77 algorithm with windowed Huffman coding. Post-training data is introduced from the start of mid-training rather than reserved for the final stages, combining long-context documents with instruction-following, reasoning and agentic trajectories; a large-scale task-synthesis pipeline grounded in task taxonomies produced more than 100 million unique tasks.

IFM traces the approach back to what it calls introducing the fully open principle in a 2023 LLM360 paper, saying it has released open models every year since, extending the commitment to larger scales, stronger capabilities and, with K2 Horizon, the complete lifecycle through agentic post-training.

Key facts

  • IFM released K2 Horizon, a connected fleet of six open models (375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B), publishing the full training pipeline, checkpoints, data recipes, code and logs, for every one of them rather than only the final weights.
  • The 0.9B, 3.7B and 7B models set new state-of-the-art results for their size classes; K2 Horizon 0.9B scored above 48 on AIME 2026 while staying small enough for watches and glasses under quantization.
  • The 36B-A4B model introduces MoVA (Mixture-of-Value Attention), extending mixture-of-experts sparsity from feed-forward layers into the attention mechanism, and needs only about 4 billion active parameters per token to perform only slightly below the dense 32B model.
  • The largest model, 375B-A23B, holds 375 billion total parameters but activates about 23 billion per token; IFM says it ranks among the top models below 400 billion parameters on general, reasoning, coding and agentic evaluations.
  • Models and code ship under the Apache 2.0 license, with datasets released under their own applicable licenses such as ODC-BY; each model was pretrained on about 20 trillion tokens, with roughly 10 trillion synthetic tokens used across pretraining overall.

Why it matters

K2 Horizon is not just another open-weight release. IFM is publishing the training lifecycle itself, checkpoints, data recipes, code and logs, for a fleet whose smallest and largest members both claim leading results in their size classes. That combination is the point: IFM frames most open releases as a tradeoff between a transparent model that falls behind the capability frontier and has little value even for research, and a powerful model shipped only as final weights that can be run but not understood. K2 Horizon is IFM's attempt to avoid that tradeoff, and it extends a yearly cadence of open releases that IFM traces back to a 2023 LLM360 paper. The MoVA architecture inside the 36B-A4B model is also a genuine mechanism change, moving mixture-of-experts style sparsity out of the feed-forward layers, where it normally sits, and into attention itself.

Who it affects

Researchers get a rare, complete record, intermediate checkpoints and fine-grained logs across pretraining, reasoning and agentic post-training, to study how tool use, planning and agentic capability actually emerge during training, and to reproduce or adapt the methods rather than starting from an opaque final checkpoint. Developers building for constrained hardware get three small models, 0.9B for watches and glasses, 3.7B and 7B for phones and other on-device applications, that IFM says deliver state-of-the-art or near-frontier results for their size. Teams running local workstations or efficient serving get the 32B dense model and the 36B-A4B sparse model as a matched pair for comparing architectures under identical conditions. Enterprises with demanding reasoning, coding or long-horizon agentic workloads are the stated target for the 375B-A23B model.

How to use it

Models and code are released under the Apache 2.0 license. Datasets are released under their own applicable licenses, such as ODC-BY, where redistribution is legally possible; where it is not, IFM discloses how the data was constructed and mixed instead of releasing the data itself. All six models support quantization, which is what lets the 0.9B model target watches and glasses specifically, while the 3.7B and 7B models are pitched at phones and other on-device applications, the 32B and 36B-A4B models at local workstations and efficient serving, and the 375B-A23B model at enterprise deployment. IFM's announcement does not give pricing, API availability, or download links.

How solid is it

Every figure here, the AIME 2026 score above 48 for the 0.9B model, the state-of-the-art claims for the 0.9B, 3.7B and 7B models, and the top-model-class rankings for 32B and 375B-A23B, comes from IFM's own announcement rather than an independent benchmark or a third party. IFM names no specific rival model in any of its comparisons: phrases like "some much larger models" and "top models below 400 billion parameters" do not identify which models were actually compared against. For SWE-bench and BrowseComp, IFM states only that the 3.7B and 7B models show strong performance, without publishing the scores themselves. No individual researcher or author is named anywhere in the post; it is written entirely in the first person plural as IFM.

Risks and caveats

IFM says complex tasks that require extensive exploration and repeated recovery, the kind TerminalBench is built to test, remain difficult for the smallest models even as the fleet overall pushes the performance boundary at every scale, so the smaller models' agentic capability has a real ceiling. The benchmark comparisons also name no specific competing model, and key results for SWE-bench and BrowseComp are described only in qualitative terms rather than with scores, so outside evaluators have not yet had a chance to confirm IFM's performance claims against named alternatives. The training data is also heavily synthetic: IFM's own pipelines generated about 10 trillion synthetic tokens used across pretraining, and more than 100 million of the tasks used in post-training were synthesized as well.