The Intelligence
Frontier reasoning, built for builders
Every agent in your company thinks with K2V3 — a frontier Mixture-of-Experts reasoning model from the lineage of one of the world's leading foundation-model labs. Scientific rigor. Blazing speed. Purpose-built for scaleups.
Reasoning lineage from
Born in a research university
Not a startup weekend hack. The intelligence behind your company comes from the labs building sovereign, open frontier models.
The world's first AI university
Founded in 2019 in Abu Dhabi, Mohamed bin Zayed University of Artificial Intelligence is the world's first university dedicated entirely to artificial intelligence. It is led by President Professor Eric Xing, formerly of Carnegie Mellon University — a research institution built to push the frontier of what machines can reason about.
Foundation models, built from scratch
IFM builds some of the world's most powerful foundation models — open, fast, and focused on solving real-world problems. Headquartered in Abu Dhabi with research hubs in Silicon Valley and Paris, IFM is the lab behind the K2 reasoning line: models engineered for scientific rigor and sovereign, end-to-end control.
The K2 lineage
Three generations of frontier reasoning
K2-Think
32B · dense · Sept 2025The reasoning breakthrough: a compact model trained with long chain-of-thought and Reinforcement Learning with Verifiable Rewards — reasoning you can check, not just trust. Demonstrated ~2,000 tokens/sec on Cerebras wafer-scale hardware.
K2-Think V2
70B · dense · Jan 2026A fully sovereign, next-generation reasoning system built on IFM's own from-scratch frontier base model — open end-to-end, from weights to training recipe.
K2V3
~375B · Mixture-of-Experts · powers SIABInside SIABThe engine inside Scaleup-in-a-Box. A frontier-scale sparse Mixture-of-Experts reasoning model with native tool-calling, served on our own vLLM infrastructure. Not yet publicly released — you meet it here first, running your company.
Mixture of Experts
A frontier brain that only wakes the experts it needs
K2V3 is a sparse Mixture-of-Experts model in the 375-billion-parameter class. Instead of firing every parameter for every token, a learned router sends each problem to a small set of specialized experts.
The result is the defining trade every scaleup wants: the knowledge of a giant model at the speed and cost of a small one. Frontier capability, without frontier latency.
Only the experts a task needs activate per token — sparse, fast, and cheap.
Why it matters for scaleups
Scientific rigor, at the speed of entrepreneurship
A scientific foundation
Reinforcement Learning with Verifiable Rewards means the model is optimized against provably-correct answers — math that checks out, code that runs, logic that holds. That structural, scientific rigor is exactly what a scaleup needs when decisions compound: a financial model, a go-to-market plan, an architecture call. Reasoning you can audit.
Lightning fast
A Mixture-of-Experts model activates only the handful of experts a task actually needs — so you get the knowledge of a 375-billion-parameter system at the speed and cost of a far smaller one. The K2 line was built for wafer-scale throughput. Your team moves at the speed of your ambition, not your patience.
A team of experts, by design
Mixture-of-Experts is the perfect architecture for a company of specialists. Under the hood, K2V3 routes each problem to the experts best suited to solve it — the same way your CFO, CMO, and CTO each own their domain. The model's structure mirrors the org it powers.
Native tool-calling
K2V3 orchestrates directly — spawning sub-agents, reading and writing files, calling tools and MCPs without a scripted shim. That is what lets ten agents behave like a real organization instead of a chatbot with a script.
Sovereign & self-hosted
The K2 line is open and sovereign end-to-end, and we run K2V3 on our own infrastructure. Your company's thinking happens on hardware we control — not rented from a black box.
The culture of speed
Entrepreneurship rewards velocity. A model born in a research university but engineered for real-world speed embodies exactly that spirit — frontier science, shipped fast. It is the perfect brain for founders who refuse to wait.
The full stack
The model is the engine. SIAB is the company around it.
A frontier model is raw horsepower. Scaleup-in-a-Box adds the skills, tools, and orchestration that turn it into a working organization.
K2V3 · the reasoning engine
A ~375B-parameter Mixture-of-Experts model with native tool-calling and verifiable-reward reasoning.
89 skills
Domain playbooks — financial modeling, GTM, product specs, presentations — that turn raw reasoning into founder-grade deliverables.
MCPs & connectors
Live tools and integrations that let your agents act in the real world, not just describe it.
10-agent orchestration
A CEO that delegates to nine department heads and their sub-agents — a mixture of experts, made org.
Put a frontier brain to work
Ten agents, 89 skills, and a Mixture-of-Experts reasoning engine — the full intelligence stack, running your scaleup from day one.
MBZUAI and the Institute of Foundation Models (IFM) are the institutions behind the open K2 reasoning line; their names and logos are the property of their respective owners and are used here to credit the research lineage. K2V3 is an evolution of that line, run by Scaleup-in-a-Box on its own infrastructure, and is not a publicly released model.