We create world-class foundation models that redefine what AI can do.
Sunnyvale • Abu Dhabi • Paris
- Tweets 174
- Following 10
- Followers 3,696
- Likes 53
Pinned Tweet
Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters.
- Frontier performance: Across coding and agentic tasks, K2 Horizon delivers top-tier performance in every size class—with the 0.9B, 3.7B and 7B models setting new state of the art at their respective scales.
- Radical openness: K2 Horizon represents the largest fully open-source model launch in AI history. The fully open code, training data and recipes are a significant step forward in transparency.
Launch page: ifm.ai/k2/
Tech blog: ifm.ai/blog/k2
Hugging Face: huggingface.co/collections/I…
173
362
1,984
1,258,586
That's a wrap on our K2 Horizon hackathon at the University of Michigan.
Congrats to K2 Ontology Generator, which turns raw CSV/JSON data and a user-defined goal into a task-specific knowledge graph, powered by K2 reasoning.
Learn more: ifm.ai/k2/
#K2Horizon
13
718
One minute from HackCMU: heads down, keyboards going, K2 Horizon running builds through the night.
This is what a glass box, not a black box, makes possible — full access to weights, code, and training data, so nothing stood between an idea and a working build.
Stay tuned, and come join us at future IFM events.
#HackCMU #K2Horizon
1
1
14
706
How big is this model? At @ICML in July we previewed K2-Horizon-3.7B and 7B and asked attendees to guess their size from watching them conduct real-work and coding tasks. Average answers, 89.9B and 62.6B.
Both off by an order of magnitude.
Density of intelligence matters more than parameter count now. The need for capable models at smaller sizes, lower cost and higher speed was overlooked. K2 Horizon is built around it.
Try them:
huggingface.co/IFM/K2-Horizo…; huggingface.co/IFM/K2-Horizo…
3
6
43
3,488
Closest guess @TianyuChen4869 won a Mac mini, which will happily run the our 7B and 3.7B models. Thanks to everyone who played.
2
240
K2-Horizon-36B-A4B scores 25 on the Artificial Analysis Intelligence Index, matching models with over 20× the total parameters while using 4B active parameters per token.
These capabilities come from our new architecture MoVA (Mixture-of-Value Attention), which incorporates MoE-based sparsity into the compute of value vectors in multi-head attention.
It opens a second axis for scaling sparsity in an LLM, beyond MoE in the FFN module. Importantly, MoVA enjoys the following advantages:
• Simple and compatible with efficient attention algorithms, such as flash attention, GQA, and sparse attention
• No additional KV cache cost comparing to standard GQA
K2-Horizon-36B-A4B available at: huggingface.co/IFM/K2-Horizo…
73
105
490
165,960
Grouped-query attention (GQA) reduces KV-cache memory by sharing keys and values across attention heads. More KV groups give more independent representations, at the cost of larger cache.
MoVA adds capacity within the existing groups instead. Sparse value experts introduce additional parameters and nonlinearity into the value computation, while the value vectors keep their original dimensions. Each token selects 4 of 64 value experts, activating just 6.25% of the value-expert pool. The value matrix gains modeling capacity at no additional KV-cache cost.
4
6
31
1,384
Training MoVA produces fewer and smaller extreme gradient spikes in a 36B-A4B comparison against standard MoE, with gated attention disabled in both. Across 45,188 matched training steps after warmup and early restarts, MoVA triggered gradient clipping 26 times against 52 for standard MoE, a 50% reduction. Its 99.9th-percentile gradient norm was 0.42 against 1.24, and its largest spike reached 64.9 against 190.9.
5
26
954
Michigan builders: K2 Horizon is coming to the University of Michigan this Saturday. ⛰️
Build with IFM’s open 375B-A23B flagship in the “Best Use of K2 Horizon” challenge.
🏆 100M tokens for every winning team member
📅 Sept. 19 | 9 AM–9 PM
📍 East Hall 1324
#K2Horizon #OpenSource
6
2
19
1,131
Today’s LLMs still write like typewriters: one token at a time. This sequential process creates a hard inference bottleneck.
We're introducing Uno, a diffusion-augmented LLM that delivers autoregressive quality at diffusion speed. It’s a lossless speedup method that accelerates generation without degrading response quality.
With Uno, K2-Horizon-7B outperforms state-of-the-art diffusion methods in both quality and throughput, delivering up to a 2.2× speedup with no loss in quality.
Paper: arxiv.org/abs/2609.04010
Model available at: huggingface.co/IFM/K2-Horizo…
91
178
1,025
237,129
Uno keeps the causal LLM architecture and adds a plug-and-play diffusion adapter alongside the autoregressive weights. AR weights for quality, the adapter enables multiple-token generation in parallel for speed.
The adapter is small and cheap to train: under 4% of the AR model's parameters, and fewer than 0.01% as many training tokens as the backbone.
1
1
30
2,531
The speedup is lossless by construction. Our Ψ-Spec sampler performs provably lossless multi-token prediction from the AR distribution, so what Uno returns is the AR model's own output.
Unlike speculative decoding, Uno:
• Does not require a separately trained draft model
• Adds fewer parameters
• Uses less memory during inference
1
28
2,003
600+ CMU students. 39 teams. 24 hours of building with #K2Horizon. 🚀
At #HackCMU 2026, teams created everything from 3D design tools to food-planning and social apps.
Congratulations to K2 Cascade, winner of the “Best Use of K2 Horizon” track! 🏆
Thank you, @nebiusai, for powering the compute behind our HackCMU sponsor track.
3
1
21
1,708
MAGENTA, from our researchers, keeps failures inside the solving loop. Failures are how science moves forward, and that’s aligned with our mission: AI that contributes to science itself.
The loop works. With MAGENTA, K2-Horizon-7B solves all six IMO 2026 problems, the smallest model reported to do so.
Try K2-Horizon-7B: huggingface.co/IFM/K2-Horizo…
🚀🚀 Introducing MAGENTA 🚀🚀
MAGENTA is the first agentic pipeline that brings natural-language reasoning and formal verification into a feedback loop, achieving perfect scores on AIME 2025, AIME 2026, and HMMT (February 2026) with various reasoning models.
🔥 MAGENTA achieved 100% accuracy across 93 problems from AIME 2025, AIME 2026, and HMMT February 2026 using K2-Horizon models of different sizes.
🔥 With the compact K2-Horizon-7B, MAGENTA solved all six IMO 2026 problems. The solutions were assessed using automated grading with reference answers and were separately reviewed by IMO medalists.
🔥 To the best of our knowledge, K2-Horizon-7B paired with MAGENTA is the smallest reported reasoner to solve all six IMO 2026 problems. This result suggests that verification and iterative refinement can allow compact models to achieve performance normally associated with much larger systems.
3
13
1,296
K2-Horizon-7B scores 21 on Artificial Analysis Intelligence Index, punching above its weights.
It scores 50% above the next best model under 10B, beating Qwen3.6-35B-A3B and closely matching Qwen3.6-27B.
• Try it: huggingface.co/IFM/K2-Horizo…
• 8-bit quantization: huggingface.co/IFM/K2-Horizo…
• Uno, the speed-up diffusion-augmented version: huggingface.co/IFM/K2-Horizo…
12
25
274
16,811
From 3.7B to 36B, whatever size a developer needs, the best model in that band is a K2-Horizon: the smartest at its size, the smallest at its level of intelligence, and interchangeable with the rest of the fleet.
K2-Horizon-MoVA-36B-A4B, K2-Horizon-7B, and K2-Horizon-3.7B all sit on the Pareto line, delivering the best cost per intelligence across model sizes. Built as one fleet, they share the same vocabulary and chat templates, so teams can choose the right model for each workload without changing how they build.
Specially:
• K2-Horizon-MoVA-36B-A4B: Matches dense 27B models on 4B active parameters per token.
• K2-Horizon-7B: Beats dense open models three times its size.
• K2-Horizon-3.7B: Best model under 4B
All available on Hugging Face under Apache 2.0:
• K2-Horizon-MoVA-36B-A4B: huggingface.co/IFM/K2-Horizo…
• K2-Horizon-7B: huggingface.co/IFM/K2-Horizo…
• K2-Horizon-3.7B: huggingface.co/IFM/K2-Horizo…
17
31
259
14,540
K2-Horizon-7B is also available at higher speed with the diffusion adapter Uno. K2-Horizon-7B-Uno delivers up to a 2.2× speedup with no loss in quality.
Try it now: huggingface.co/IFM/K2-Horizo…
2
9
1,262
Intelligence is too important to buy as a service and too fundamental not to own as a capability.
K2 Horizon provides not only weights entities can hold, but a foundation they can build their own intelligence on without external dependencies.
K2 Horizon is here to put that ownership in your hands: run it, inspect it, retrain it, keep it.
4
5
17
1,236