Multiverse unveils LittleLamb open-source AI models

Multiverse Computing has released LittleLamb, a family of open-source language models. The launch adds three compact models for edge, mobile and offline use.

The new line includes LittleLamb 0.3B, LittleLamb 0.3B Tool-Calling and LittleLamb 0.3B Mobile. All three are based on Qwen3-0.6B and have been compressed to about half the size of the underlying architecture using Multiverse's CompactifAI technology.

The release comes as debate over European AI sovereignty has focused heavily on access to cloud infrastructure, data centres and computing resources. Multiverse argues that another issue lies closer to the point of use: whether AI models can run locally in constrained environments instead of depending on remote cloud systems.

That is relevant for organisations using AI in mobile devices, industrial systems and defence-related settings, where low latency, offline access or tighter control over deployment may matter. Compact models also offer an alternative for developers working within memory, battery and bandwidth limits.

Each model has a footprint of about 0.3 billion parameters and supports English and Spanish. They also offer two inference modes: one focused on more detailed reasoning for tasks such as mathematics and multi-step problem-solving, and another aimed at faster responses in general dialogue.

Multiverse has positioned the three models for different tasks. The standard LittleLamb 0.3B is intended for general conversational AI, question answering, reasoning and virtual assistants on edge or on-device systems.

The Tool-Calling version is tuned for function calling, structured JSON outputs and agentic workflows. It is designed for developers who want a compact model that can interact with APIs, browsers, code execution environments and other integrated systems within automation pipelines.

The Mobile variant is packaged for inference on mobile and edge hardware. It is aimed at on-device assistants and offline applications where memory, latency and battery constraints are tighter.

Benchmark claims

Multiverse said LittleLamb 0.3B and LittleLamb 0.3B Tool-Calling outperformed the original Qwen3-0.6B model and models in the Gemma 270M class on HLE testing. It also said the compressed models improved throughput, latency, output speed and time-to-first-token, while the Mobile version improved accuracy on mobile action tasks against the same class of rival models.

More broadly, the company argues that model compression can expand the range of environments in which language models can be used. Rather than relying on larger systems hosted by hyperscale cloud providers, smaller models can be embedded closer to the application, whether in a handset, a factory device or an offline terminal.

According to Multiverse, CompactifAI uses quantum-inspired tensor network mathematics. The company says the technology can reduce model size by up to 95% with a precision loss of 2% to 3%, compared with what it describes as an industry norm of 20% to 30% accuracy loss at similar compression levels.

Multiverse, headquartered in Donostia, Spain, with offices in North America and Europe, says it serves more than 100 customers, including Iberdrola, Bosch and the Bank of Canada. The release of LittleLamb extends its catalogue of compressed open-source models as it seeks to build a position in edge-focused AI deployment.

Open-source distribution may also help the company reach developers more quickly in a market where smaller language models are becoming more competitive. Businesses and public sector users are weighing cost, control and data handling alongside raw model size, especially for deployments that do not require the full scale of larger frontier systems.

In Europe, that debate increasingly touches on sovereignty in practical rather than purely infrastructural terms. If a model still needs to call back to external cloud systems to function, local ownership of servers and data centres does not fully remove external dependence.

Chief executive Enrique Lizaso Olmos framed the launch in those terms. "The launch of LittleLamb continues our mission to make efficient AI available across every deployment environment without losing the flexibility and accuracy developers need. With CompactifAI, we've demonstrated that compression doesn't require sacrificing intelligence or capability. This model family shows that compact models can do far more than lightweight chat, and can run in environments where traditional models are simply too large or too dependent on cloud infrastructure," he said.