Horizon Lens — 14 October 2026
Liquid AI targets tool use on local hardware
Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model with 24 billion total parameters and two billion active per token. Its October 13 announcement positions it for local agents that call tools and perform multi-step tasks. The release supports serving systems including llama.cpp, MLX and vLLM; the model itself is only one component of a working agent.
The company reports strong tool-use and instruction-following results, but its own comparison also shows larger models retaining an edge on knowledge-heavy and coding benchmarks. Its quoted 184 tokens per second on an Apple M5 Max is a vendor measurement on that hardware, not a promised speed for every computer or complete agent workflow.
Analysis
The interesting trade-off is task fit. A compact active computation budget may suit repeated tool calls, while a harder coding problem may benefit from a different model. Parameter counts and decoding speed do not tell you how often a task finishes correctly or how much checking the result requires.
Action
Evaluate a few repeatable tasks on your own machine, recording completion quality and total elapsed time. Include failed tool calls and corrections in the comparison, rather than judging by response speed alone.
PAIR leaves beta and broadens local-network inference
NVIDIA says PAIR, its free, open-source personal AI router, moved out of beta on October 13. The software discovers compatible computers on a local network and distributes AI workloads across available systems. It runs on Windows, Linux and macOS, making the announcement relevant beyond the company’s upcoming RTX Spark laptops.
The official release adds llama.cpp support, direct model downloads from Hugging Face, configurable inference-engine settings and RTX Spark support. Separately, NVIDIA says its Nemotron 3.5 Bolt 74B model becomes available on October 15. The router release and the model’s future availability are different milestones; they should not be collapsed into a single claim that everything is ready today.
Analysis
This makes a local network another place to allocate work, rather than assuming every request belongs on the computer in front of you. The practical benefit depends on compatible devices and the jobs being run. Discovery and routing support alone do not demonstrate a speed improvement for a particular household or office.
Action
If trying networked inference, first establish one reliable workload on one machine. Then compare the routed version for completion time and reliability, keeping track of which computer actually handles each request.
HomeView moves from preview to announced product
Apple has now announced HomeView, following the preview covered earlier this week. The Verge’s October 13 first look describes a touchscreen above a speaker base for video calls, photos and information through Siri AI. Preorders are open, with shipping scheduled for November 13; an announcement and an order page are not the same as widespread household use.
The report lists the UK among eight launch countries. It describes HomeView as an Apple Home hub, Matter controller and Thread Border Router, able to coordinate compatible smart-home equipment. Siri interactions can draw on messages, calendars and email. Our source capture covers the publicly visible introduction, so this is a launch summary rather than a full product assessment.
Analysis
The distinction worth testing is between personal assistance and a shared household screen. Convenient access to contextual information raises practical questions about who sees what and whose instructions are acted on. Compatibility labels help describe the connections, but they do not settle how well a family’s daily routines will work.
Action
Before planning a setup, list the household members, devices and routines it must support. Check the published compatibility and account controls, then look for evidence from real use after shipping begins.
A retinal implant milestone, with clear limits
Science Corporation says its first patient with Stargardt disease received the PRIMA retinal implant in May at Sydney Eye Hospital. CEO Max Hodak discussed the milestone at TechCrunch Disrupt on October 13. The company reports that the patient saw a letter in the first rehabilitation session and read a word in the second.
The system pairs an implant at the back of the eye with camera-equipped glasses. Hodak described a small, high-contrast black-and-white patch in central vision, allowing a few letters to be seen at a time, rather than a wide field of colour vision. This report concerns one company-described patient milestone, not proof of general effectiveness for the disease.
Analysis
The engineering progress is more understandable when the output is described precisely. Recognising letters can be meaningful without equating to restored normal sight. The May procedure and October announcement also belong on separate dates: fresh reporting does not mean the intervention happened this week.
Action
When following implant research, separate reported individual progress, the limits of the device and trial evidence across patients. Look for published outcomes and follow-up before drawing broader conclusions from a demonstration or company account.
Remote hiring remains a route for alleged insider access
US prosecutors say they have disrupted a remote-worker operation involving alleged North Korean operatives and American facilitators, TechCrunch reports. According to the Justice Department, authorities searched more than 40 suspected laptop farms in Texas and seized over 180 computers. The report describes arrests and allegations, not convictions establishing every claim.
Authorities believe companies shipped the machines to people hired under stolen American identities, enabling access from elsewhere. Prosecutors said at least 70 US-based companies had been infiltrated; the affected businesses were not named. The DOJ also warned that facilitators are taking on a wider range of tasks as companies improve detection.
Analysis
The operational issue is the gap between an apparently ordinary employment setup and the person actually controlling the equipment. A familiar shipping address or convincing online profile does not independently establish identity. The case illustrates why hiring checks and device access should be considered together, rather than as unrelated administrative processes.
Action
For remote onboarding, review how identity, equipment delivery and subsequent access are connected in your existing process. Investigate inconsistencies through established security channels rather than relying on nationality or a polished online profile as a shortcut.