IFM releases six Apache-licensed K2 Horizon models with training checkpoints, data recipes, logs, and code
- IFM released K2 Horizon, a six-model family under Apache 2.0: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B parameters.
- IFM says it is publishing the full training lifecycle for every model, including intermediate checkpoints, final weights, architecture, mixture compositions, configurations, fine-grained logs, evaluations, and training code; it releases datasets where redistribution permits and otherwise publishes construction and mixing recipes.
- The largest 375B-A23B model uses sparse MoE with 375 billion total parameters and about 23 billion active parameters per token, while the 36B-A4B model uses IFM's Mixture-of-Value-Attention mechanism.
- IFM reports that its 0.9B, 3.7B, and 7B models lead their size classes on mathematics, reasoning, coding, and agentic evaluations; the 0.9B model scored above 48 on AIME 2026.
- The fleet shares architecture, vocabulary, training methodology, interfaces, evaluation infrastructure, and deployment tooling, except that the 0.9B model has a smaller vocabulary; all six models support quantization.
Hacker News opinions
I'm getting model fatigue. New models arrive far faster than JavaScript frameworks did a decade ago, though at least they are easier to adopt.
This one does have a real differentiator: it claims to be fully open across the training stack.
I think models are following the CPU arc. Eventually most people will stop tracking every release and just pick what is good enough for the job.
I want fully open models, including source code, training data, processing, and data organization. Otherwise users cannot know what sits beneath a closed model or rule out social manipulation.
That only works if the training data has permissive licenses, and funding a fully open stack is the hard part.
The original link pushed me to a login page. worked for me and lists the 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B variants.
The pre-training and post-training repositories were empty when I checked. It looks like someone may have published the announcement before the release artifacts.
The benchmark comparisons are strange because the competing models change by task. I would rather see recent open models such as DS4Flash, GLM53Flash, and Qwen38 on the same suite.
The fully open stack is welcome, but the headline claims look stronger than the charts. The 32B model trails Qwen3.8 27B, Gemma4 31B is absent, and that size is the self-hosting sweet spot.
The 32B is labeled stage 1 with a final checkpoint still to come, which may explain the result. The 7B does look very good.
I will believe "radically open" when the training data ships with the weights. Until then, it is a fast demo.
I checked Hugging Face and found training data there, including 3.3 TB for code reasoning, 4.5 TB for mathematical reasoning, and 8.4 TB of pre-training behaviors. That is still likely far less data than the largest closed-model labs use.