AI’s next leap isn’t just about bigger models, but smarter ones. We've got an inside look at a breakthrough reasoning model that squeezes more performance out of fewer parameters.

It's rare for efficiency, transparency, and capability to move forward together, but that's what we're seeing. MBZUAI’s K2 Think V2 is positioning itself as both a technical standout and a philosophical one, challenging assumptions about scale, openness, and control.

We break down how its training approach cuts hallucinations, why its end-to-end openness matters in a market full of half-open claims, and what a “fully sovereign” model means for enterprises, governments, and builders navigating an AI landscape increasingly defined by trust.

Lean AI model sets new standard for efficiency

With advances in reasoning, AI models are making significant strides in efficiency and cost savings. A new model from MBZUAI’s Institute of Foundation Models is pushing the boundaries.

On Tuesday, the institute released K2 Think V2, its first fully sovereign general reasoning model. The lightweight, 70-billion-parameter model offers competitive performance compared to other models of its size and can keep up with “significantly larger” models across a number of benchmarks, Hector Liu, Head of Technology at IFM, told The Deep View.

To put it plainly, this model can do a lot more with a lot less, said Liu: “This translates into more cost-effective inference without sacrificing reasoning quality.”

The difference is in the training, said Liu. Instead of focusing on “extreme parameter scaling,” K2 Think V2 was built on the Instruct model, designed specifically for thinking and reasoning.

• Additionally, the model was trained using a two-stage system called reinforcement learning with verifiable rewards, which rewards the model when it produces verifiably correct answers. This means the model doesn’t just parrot back what a correct answer looks like, but is actually thinking about how to answer the question properly.
• The model’s developers also used disciplined dataset curation, relying only on “filtered and fully decontaminated” data focused on math, coding and STEM.

This training method resulted in “industry-leading” low hallucination rates, said Liu, establishing a “foundation of truth by carefully curating our training datasets and strictly validating them for correctness.”

So what can K2 Think V2 actually do? Basically, this model is very, very good at math. Compared to previous model iterations, the system saw substantial performance improvements on the American Invitational Mathematics Examination, Harvard–MIT Mathematics Tournament, the diamond tier of the Graduate-level Google-Proof Question Answering, and the Instruction-Following Benchmark.

In practice, this model excels at problems that require long chain-of-thought reasoning and step-by-step logic, without going haywire over long contexts. That makes it a good fit for regulated environments like research institutions, governments, or enterprises — areas where teams need strong reasoning capabilities, full transparency and control over deployment.

“We conduct rigorous experiments when introducing new capabilities to ensure that expanding the model's skill set never comes at the cost of its reliability,” Liu said.

K2 Think V2 ranks among AI’s most open models

Open models have gained significant traction over the past year, and now the Institute of Foundation Models at MBZUAI is taking openness to a whole new level.

K2 Think V2 is one of the few models available that’s open “end-to-end,” offering transparency through every part of the stack. This means that users get access to its weights, training data, code, intermediate checkpoints and evaluation tooling, IFM’s Liu told The Deep View.

Though popular open models like DeepSeek, Alibaba’s Qwen, and OpenAI’s GPT-OSS tout their openness, these models are all open-weight rather than open-source. This means that only their trained parameters are publicly available.

• Comparatively, K2 Think V2’s complete openness allows the developer community the ability to know it inside and out, facilitating easier inspection, reproduction and innovation for developers, and offering a “level of transparency that other leading open models do not offer,” said Liu.
• This allows for genuinely independent evaluations, as it doesn’t rely on proprietary datasets and hidden pipelines that might artificially boost results.

In practice, this means that organizations know exactly what they’re getting when they deploy, audit and build on K2 Think V2. That’s because this model’s reasoning outputs can be traced back to “concrete training choices rather than black-box effects,” said Liu.

IFM’s latest model comes amid an inflection point for open models. With the cost to deploy AI rising, many developers and enterprises are seeking alternatives to major model providers. Some are turning to open-source technologies to more affordably keep up with the rapidly changing AI landscape. Still, because many only open their weights, users aren’t always aware of what went into developing these models.

K2 Think V2, meanwhile, may offer a far clearer picture than its competitors. “The main benefit is independence and credibility,” Liu said.

Why is it the 'first fully sovereign' model?

Alongside its focus on efficiency and open-source, the other key aspect of K2 Think V2 is that it's what IFM calls the first "fully sovereign" reasoning model.

That's because IFM offers complete end-to-end transparency for the model, with no proprietary sources or external pipelines. As a result, any developers drafting off of this model to build their own AI systems can have confidence in the underlying code and can use their own data sources to meet whatever standards are required by regulatory bodies.

Liu also pointed out to The Deep View that "IFM can continue to independently evolve and refresh the model as new architectures, data, and RL methods are developed. Sovereignty here is first about builder control and independence."

That will be music to the ears of IT leaders and software engineers in highly regulated industries such as healthcare and banking. It makes K2 Think V2 "well-suited for sovereign or regulated environments, like research institutions, government, or enterprises, where teams need strong reasoning performance but also require full transparency, evaluation credibility, and control over deployment and adaptation," said Liu.

In this case, the model's open-source nature is closely related to the fully sovereign claim. So these openness factors play a key role in the sovereign model premise:

• The model is open from pre-training through post-training
• Only uses IFM-curated or synthesized data
• Has explicit decontamination from downstream evaluations

What IFM pulled off is a combination of strong reasoning performance and full transparency. That remains rare for a model at this scale, even though a 70-billion-parameter model is small, less than a tenth the size of today's flagship models from OpenAI and Google that clear a trillion parameters.

When we pushed Liu on the importance of sovereignty in this model, he asserted, "For developers and organizations, this enables practical sovereignty: the ability to deploy the model on their own infrastructure, fine-tune it with domain-specific data, and apply their own governance and compliance standards. The model enables sovereignty by design; how it’s applied depends on the user’s needs."