LiquidAI has introduced LFM2.5-2.6B, a compact on-device model for tool use and multi-step workflows, while Thinking Machines has unveiled Inkling, an open multimodal model supporting image, text and audio inputs with a one-million-token context window.
LiquidAI and Thinking Machines have published separate Hugging Face launch posts describing new open-model efforts aimed at distinct AI workloads: compact local automation in one case, and long-context multimodal processing in the other.
LiquidAI’s post introduces LFM2.5-2.6B, which the company describes as an on-device agent model built for tool use and multi-step workflows on everyday hardware. Thinking Machines’ post presents Inkling as an open multimodal model that can accept images, text and audio, with support for a context window of up to one million tokens.
The announcements do not offer a direct comparison. Instead, they illustrate different technical priorities that developers may weigh when selecting or testing open models.
According to LiquidAI’s Hugging Face launch post, LFM2.5-2.6B is intended to run locally while handling workflows that involve multiple actions and external tools. Such workflows can include calling functions, retrieving information or moving through structured steps rather than producing a single standalone response.
The model’s 2.6-billion-parameter scale is central to its positioning. Smaller models can be more practical for deployment on local or edge-oriented hardware than much larger systems, particularly where developers want more control over latency, connectivity or data handling. LiquidAI describes the model as designed for everyday hardware, although its launch post does not provide detailed hardware specifications, comparative benchmarks or independently verified performance figures.
Tool use and multi-step execution are relevant for applications that need a model to participate in software processes. A local model in this category could be useful where an application must invoke defined tools or complete repeatable task sequences without making every interaction dependent on a remote service.
Thinking Machines takes a different approach with Inkling. Its Hugging Face post describes the model as open and multimodal, accepting image, text and audio inputs. That combination could allow developers to build systems that work across written documents, visual material and recorded sound through one model interface.
The company also says Inkling supports up to one million tokens of context. In practical terms, a large context window can enable a model to receive much larger bodies of material in a single interaction, potentially including long document collections or extensive mixed-media inputs.
However, the launch post does not detail evaluation methods, serving requirements, latency, memory use or performance trade-offs at the stated maximum context length. Developers considering the model would therefore need to assess those operational factors for their own workloads.
The two releases point to separate directions in open-model development. LiquidAI is emphasizing compact deployment and local, tool-using task execution. Thinking Machines is highlighting the ability to process several input modalities alongside a very large amount of context.
Neither source establishes that one approach is generally superior. A local automation application may prioritize model size, responsiveness and tool-use behavior, while a media-analysis or document-heavy system may place more value on multimodal support and context capacity.
The Hugging Face announcements provide an early outline of each project’s intended use. Independent testing, deployment documentation and workload-specific evaluations will be needed to determine how the models perform in production settings.
LiquidAI’s post introduces LFM2.5 2.6B , which the company describes as an on device agent model built for tool use and multi step workflows on everyday hardware.
Thinking Machines’ post presents Inkling as an open multimodal model that can accept images, text and audio, with support for a context window of up to one million tokens.
Instead, they illustrate different technical priorities that developers may weigh when selecting or testing open models.
Continue reading
