Local AI on a €2,960 box

A structured, pre-registered benchmark of what one 128 GB mini-PC can actually do - chat, coding, transcription, speech, images, documents - measured, sourced, and clearly framed.

~93tokens/sec chat - faster than you read
28×real-time meeting transcription
285-730×cheaper per word than frontier cloud AI
100%on-device - nothing leaves the building

Read the one-page briefing Full report Charts

Alastair McDermott holding the GMKtec EVO-X2 mini-PC

The actual box - Alastair McDermott with the GMKtec EVO-X2 this runs on. Watch the 2-minute intro on LinkedIn →

🔊 Listen to this pagenarrated on the box - script by the local 30B model, voiced by Piper TTS (how?)

Bottom line: capable AI on your own desk, with no data leaving the building - fast enough for everyday work, hundreds of times cheaper per word than a top cloud AI, and clear about where it doesn't match frontier models. Everything here is the measured evidence; every number traces to a raw file. GMKtec EVO-X2 · Ryzen AI Max+ 395 · 128GB · llama.cpp/Vulkan + ROCm 7.2.4 · benchmarked 2026-07-03→06.

See it in action

🎬 90-second explainer video

A narrated chart walkthrough - script by the local model, voiced by Piper, assembled with ffmpeg, all on the box.

🎙️ Two-minute audio overview

A two-voice conversation, voiced on the box by Chatterbox TTS. Good, but not cloud-studio grade - see the audio quality verdict. Hear the earlier Piper version.

Both were generated by the box itself - how, and who did what.

Cost per 1M output tokens - local vs cloud (log scale)
Cost per 1M output words - local vs cloud, log scale. See all five interactive charts →

What would it save you?

A rough guide, and only if your work fits what the box does well (see the limits). Assumes the box at €2,960 plus about €8 a month electricity, over three years. It does not price in data-sovereignty value, which for many buyers is the whole point.

Reports

~/reportsExecutive BriefingFor a decision-maker: what it does, who it's for, the limits, the cost - one page.→ read ~/reportsCanonical Benchmark ReportThe source of truth - every measurement with error bars, findings (F1-F23), glossary, FAQ.→ read ~/reportsCapability ReportThe narrative write-up for a business-leader reader.→ read

Findings & evidence

~/findingsFindings register (F1-F23)One durable, evidence-backed finding each.→ read ~/findingsExperiments E1-E18Pre-registrations and results, incl. Vulkan-vs-ROCm.→ read ~/findingsFrontier quality A/BLocal vs gpt-5.5 on the same suite (2026-07-06).→ read ~/findingsPhase-A evidence logEvery gate, bench, and prediction, dated.→ read ~/findingsProvenance manifestRepo, revision, size, SHA256 per model.→ read

Capabilities & visuals

~/chartsCharts (interactive)Depth, MoE-vs-dense, cost, concurrency, Vulkan-vs-ROCm.→ view ~/chartsImage gallery (50)FLUX images across 10 style families, generated locally.→ view ~/chartsNon-LLM probesSpeech-to-text, TTS, image generation, upscaler.→ read ~/chartsImage quality verdictCalibrated assessment of the local gallery.→ read ~/chartsAudio quality verdictHow local TTS was chosen by ear (Piper to Chatterbox), and where it trails cloud.→ read ~/chartsModel surveyPer-model corridor-rule checks; quant ladder.→ read ~/chartsQuality evalsDeterministic eval matrix and resolutions.→ read ~/chartsHow this site was madeWho did what: the box's AI ran the narration and images; Claude Code + a human built the report and site.→ read

See if it fits your practice

A 25-minute call, no obligation - we'll talk through whether an on-prem box like this makes sense for the work you do.

Book a call

Raw data & sources

Self-contained - no separate data download needed. data/raw/ holds every raw llama-bench table and serve_bench JSON behind the numbers; data/report-sources/ holds the editable Markdown of every report; data/eval-pilot/ and data/eval-suites/ hold the eval outputs and definitions. Full annotated index: data index.