MBZUAI Releases K2 Horizon, the Largest Fully Open-Source AI Fleet Ever

The Institute of Foundation Models, an MBZUAI-backed lab with offices in Abu Dhabi, Silicon Valley, and Paris, has released what it claims is the largest fully open-source model release in AI history. K2 Horizon is a fleet of six foundation models spanning three orders of magnitude in scale, each shipping with weights, source code, training data, and the recipes used to build them.

That last part is what separates K2 Horizon from most of what the industry calls open source. Meta, Mistral, and others release weights; IFM is publishing the full pipeline, which means researchers can audit training data for contamination or licensing issues, reproduce ablations to verify benchmark claims, and fork the training recipe to build domain-specific variants without reverse-engineering hyperparameters.

Six sizes, one architecture

The six models share a core architecture, vocabulary, training methodology, and deployment tooling, with dynamic routing directing tasks to the most cost-effective model in the family. Prototype on the 0.9B and scale to the 375B flagship without rewriting your deployment code.

IFM positions the 7B as the strongest model under 10B parameters and the 3.7B as the strongest under 4B, with both claiming state-of-the-art results on reasoning, math, coding, and agentic benchmarks. The 3.7B is also explicitly pitched as fine-tuning-friendly, suggesting IFM expects heavy derivative work at that scale.

Diffusion distillation and MoVA attention

Standard autoregressive models generate one token at a time, which sets a hard latency floor. IFM describes a diffusion distillation technique that generates blocks of tokens in parallel, claiming roughly 3x throughput gains without degrading quality. Diffusion-based text generation has been an active research area, with projects like Mercury and LLaDA exploring the space; distilling it into a production fleet is a meaningful engineering step, though IFM has not specified the hardware or context-length conditions behind the 3x figure.

The second architectural move is Mixture of Value Attention, or MoVA. Standard Mixture-of-Experts routes tokens to different feed-forward experts. MoVA applies the same sparsity idea to the attention mechanism's value projections, aiming to improve reasoning without adding compute. The 36B-A4B model activates only 4B parameters per token while reportedly outperforming larger dense models, which is what makes it viable on a workstation-class GPU.

What Apache 2.0 means for your use case

All six models are available now on Hugging Face, with serving via vLLM and SGLang and API access through Compass, Cerebras, AWS, and Nebius. The Hugging Face collection is the fastest starting point. A minimal vLLM setup for the 7B:

from vllm import LLM, SamplingParams
llm = LLM(model="IFM/k2-horizon-7b")
params = SamplingParams(temperature=0.7, max_tokens=512)
out = llm.generate(["Explain MoVA attention in one paragraph."], params)
print(out[0].outputs[0].text)

Everything ships under Apache 2.0, which permits commercial use with no revenue thresholds or acceptable-use carve-outs. That is a cleaner licensing story than Llama's community license, and it matters for teams deploying in regulated environments where legal review of model licenses is mandatory.

Caveats before you commit

A few things worth checking before building on these models:

• Independent benchmark verification is pending. State-of-the-art claims at each scale need scrutiny against current competitors like Qwen3 and the latest Gemma releases.
• The 3x diffusion distillation speedup lacks a specified hardware or context-length baseline, so real-world gains will vary.
• The 375B flagship is benchmarked against open-weight models, not closed frontier systems like GPT-5 or Claude.
• MoVA is a new architecture, and tooling support outside vLLM and SGLang may lag.

The 32B dense model fits on a single high-end GPU, and the 36B-A4B is designed for cost-sensitive local deployments where activated-parameter count drives latency. Teams targeting consumer hardware will find the 0.9B and 7B cover most realistic on-device use cases.

Geography and the open-source bar

IFM founder Eric Xing framed the release around reproducibility and collective progress, arguing that meaningful AI advancement requires the ability to examine and build on the underlying technology rather than consume it through an API. Whether the benchmarks hold up under independent testing will determine how widely the field adopts K2 Horizon, but publishing weights, data, and training recipes together raises the bar for what open source in AI should mean.