Topic
Running machine learning models on local devices like phones and laptops instead of the cloud, using quantization and optimized runtimes to fit hardware limits.
Supersonic Labs' Julia 1, a 144.3M-parameter model trained for about $104, picks labels from a caller's list on an ordinary CPU. On its own 72-label Banking77 pilot it scored 64 of 100, which shows where the 20-option cap keeps it from replacing an LLM classifier.
Reality
Liquid AI shipped a 279.5-million-parameter draft model for LFM2.5-VL-3B on September 24, reporting 2.30x to 3.13x faster decoding on an Apple M5 Max. Image encoding and prompt prefill are unchanged, and the company's tests did not cover quantized deployments.
Reality
GLM-5.3-Flash leads agentic terminal work, DeepSeek V4 Flash is billed as the cheapest per token, and a 2.52B MiniCPM5-2B runs locally under Apache 2.0. The comparison flags most of those numbers as vendor-reported.
Reality
kev packs a document and every typed question into one sequence and reads them in a single forward pass on a Mac. Its author scored four checkpoints against the real Jev on frozen items and published the reads that missed a pre-declared gate.
Reality
The ONNX export alone ran 1.58x faster than eager PyTorch with bit-identical accuracy. The int8 pass after it cost 8% in latency and bought a 64 MB download, small enough to serve as a static page with no backend.
Reality
The opencode-agent-memory plugin gives an OpenCode agent scoped Markdown blocks it rewrites through three tools, plus an opt-in journal it can search locally but never revise. The maintainer documents the failure modes himself.
Reality
Mininglamp published NavEval scores for its own model on its own benchmark. Across the three entries, the spread tracks how each stack reads a page. It is not a case of specialists beating frontier models.
Reality
Nine local VLM configurations read the same 23 handwritten clinical pages against 221 hand-curated gold values, and accuracy on the fields accepted without human review rose from 75% to 96% once a second model started checking the first.
Reality
Microsoft's local runtime handles model download, hardware detection and execution provider selection behind a .dll, .so or .dylib. An app that adopts it owns the download, the cache and the unload.
Reality
An n-gram count table was rebuilt so it could be paged from disk. The measurement went looking for a small resident hot set to hold and did not find one, and the fallback cost estimate rests on a page-read latency the project had only assumed.
Reality
Apple wants OpenAI's internal communications, the accounts that held allegedly confidential material and the employees who touched it, all before a hearing on an injunction against a device OpenAI has yet to release.
Reality
A dev.to writeup read the live pages of five background removal vendors on 8 September 2026. Remove.bg sets no duration on how long images are kept, and VideoBGRemover publishes erasure within 30 days plus a seven-day backup tail.
Reality
A wireless device worn on the animal combines flexible probes, optogenetics, motion sensing, an ultrasonic microphone and a head camera, and runs its event detection onboard so it can intervene while a mouse is still moving.
Reality
A robot walking itself off a Guangzhou line is a real assembly milestone. XPeng's own timeline still puts mass production at year-end and outside delivery in 2027, the same year Tesla now targets Optimus sales.
Reality
Portable Computer runs Perplexity's agent on 27B models on your own hardware. Paid tiers only, Linux first, Windows in September, and a 24GB VRAM floor most desktops miss.
Reality
Its August 20 bake-off priced three agents on one Nemotron port. The claim that findings transfer to the next model-chip pair is the one with no number attached.
Reality
A vendor benchmark puts wake-word at 5.3% of one core on a $15 Pi Zero 2 W. That is the bar for on-device listening now, and the free-phrase-only catch travels with it.
Reality
A dev.to writeup argues Whisper is a batch model retrofitted for streaming. If your response budget is under a second on mid-range phones, that is a build decision, not a tuning exercise.
Reality