CodeMingle AI News Report - October 15, 2026
Executive Summary
Local AI is moving from model demos toward practical workflows. Liquid AI has released a model for on-device agents, while Google DeepMind and Singapore's GovTech describe a phone-based scam detector. AMD's new rack blueprint provides infrastructure context: agent execution and model inference can have different compute needs.
Editorial date: October 15, 2026, New Zealand time. Lead announcements are dated October 13–14; AMD's October 12 release is included as infrastructure context. Performance figures are vendor reports, and planned availability is identified as such. Builder exercises and implications below are CodeMingle's analysis.
Listen to the podcast edition
Top AI News Stories
Liquid AI releases LFM2.5-24B-A2B for local agents — October 13
Liquid AI's new open-weight reasoning model has 24 billion total parameters and 2 billion active parameters per token, with a 131,072-token context window. The release includes support for llama.cpp, MLX, vLLM, SGLang, and ONNX. Liquid reports 184 tokens per second on an Apple M5 Max; that is a vendor measurement under its test conditions. Liquid AI's launch report
CodeMingle analysis: Treat a local agent prototype as a complete application. Measure the time from the user's request to a verified result, including prompt preparation, reasoning, tool calls, and retries. A fast decoder may still produce a slow workflow if it needs several attempts to finish a task.
DeepMind and GovTech put scam analysis on the phone — October 14
Google DeepMind and Singapore's GovTech describe an on-device system using a lightweight screening classifier followed by a Gemma model fine-tuned for local scam conversations. Suspicious messages receive contextual analysis; detected manipulation triggers a warning. The announcement describes local conversation processing and training on anonymised scam-report and intelligence datasets. DeepMind's announcement
CodeMingle analysis: Evaluate the first screening stage separately. If it misses a harmful message, a stronger second model never sees that case. For the user experience, make the warning understandable and give people a clear next step, such as checking a claimed sender through an independently obtained contact.
AMD separates agent execution from inference at rack scale — October 12 context
AMD introduced Helion, an open rack blueprint with CPU-only and combined CPU/GPU configurations. Its CPU-only example supports up to 32,768 cores and connects agent execution to separate inference infrastructure. AMD says partner configurations remain under development, with availability planned for November 2026. AMD's release
CodeMingle analysis: Capacity planning should begin with a trace of actual work. A workflow waiting on a browser, document parser, or external service needs a different intervention from one waiting on model inference. Separate those timings before choosing more hardware.
Technical Deep Dives (Architecture & Implementation)
Evaluate a two-stage classifier as one product
CodeMingle architecture exercise: Build a reviewed set containing ordinary messages, obvious scams, and ambiguous conversations. Run it through the complete pipeline and record where each mistake enters. Measure screening misses, second-stage false alarms, end-to-end latency, and the percentage of messages escalated.
Try changing the screening threshold and inspect both error rates and energy use. A cheaper first stage is useful only if the full product still catches the cases it was built to address. Keep a separate evaluation set when tuning thresholds so improvements reflect more than familiarity with the examples.
Budget memory beyond active parameters
CodeMingle implementation exercise: For a mixture-of-experts deployment, record checkpoint size, resident memory after loading, and peak memory at realistic context lengths. Active parameters describe which weights participate in a token's computation; they do not establish the total memory required by the runtime.
Test a short request and a long conversation on the intended machine. Keep the model version, quantization, runtime version, and generation settings with the results. Repeat with the user's normal applications open to reveal resource contention.
Developer Tools & AI Agents
Give the local agent a measurable completion rule
CodeMingle builder exercise: Pick a contained task, such as summarising a folder of sample documents into a draft. Specify the allowed input directory, output path, and review criteria. After execution, inspect whether the output exists, cites the right inputs, and satisfies the requested format.
Add a missing file and an interrupted tool call. Check whether the agent preserves progress, explains the gap, and resumes without duplicating work. Record accepted outcomes per attempt alongside response speed; this makes model comparisons relevant to users.
Hardware & Infrastructure
Trace the workflow before scaling it
CodeMingle infrastructure exercise: Divide task time into model loading, input preparation, inference, tool execution, and external-service waiting. Track concurrent tasks and the resources they hold while waiting. A queue can become congested even while an accelerator is mostly idle.
Compare a small local deployment with a hosted workflow using the same acceptance criteria. Include operational effort, failure recovery, and sustained latency in the comparison. Use announced hardware dates as planning milestones until an actual configuration is available to test.
Detailed Trend Analysis
CodeMingle editorial analysis: These announcements point toward more deliberate placement of AI work. A phone-based classifier, a local agent, and a rack-scale execution service each need evidence suited to their job.
For builders, the useful question is which part of the task benefits from a model and where that part should run. Put a completion check beside every action, measure the full path, and collect difficult cases. Those records help a team decide when to change the model, adjust the workflow, or improve the interface.
Future Outlook
Watch for independent evaluations of Liquid's model, concrete integration details for the scam detector, and AMD partner configurations ahead of the planned November availability. Compare results only after checking the task, runtime, and measurement conditions.
Today's challenge: Trace one AI workflow from input to outcome. Find its slowest stage and write one test that would expose a failure there.