MERKIUM

Models

The Kili family of models

From a compute-efficient foundation model to our frontier flagship, every Kili model is engineered for reliability, safety, and real-world performance.

Flagship · June 9, 2026

Kili 2.5

Our most capable model to date. With 230 billion parameters and a 590K-token context window, Kili 2.5 delivers frontier reasoning, agentic tool use, and deep multimodal understanding — purpose-built for the hardest problems in research and production.

Read the technical details

230B

Parameters

590K

Context window

Multimodal

Text, image & code

Agentic

Native tool use

Compare the family

Kili 2.5

June 9, 2026

Flagship · 230B

Frontier-class reasoning, agentic tool use, and deep multimodal comprehension with a 590K-token context window — built for the hardest problems in research and production.

Complex reasoning, agents, research

Kili 2.1

May 31, 2026

Refined · 70B

A targeted refinement of Kili 2 with sharper code generation, stronger mathematical reasoning, and improved multilingual performance for demanding teams.

Coding, multilingual, lower latency

Kili 2

May 1, 2026

General · 34B

A balanced, production-grade model delivering strong reasoning and instruction adherence across general-purpose workloads and tool-use scenarios.

General workloads, tool use

Kili 1

February 13, 2026

Foundation · 7B

The model that started it all. Compute-efficient and highly capable for everyday text tasks, on-device inference, and lightweight deployments.

Fast inference, on-device