Could Selma run on a solar panel?
Offline, yes, today. On one solar panel, not with the current tower. A smaller box would make it true.
Offline, yes, today. On one solar panel, not with the current tower. A smaller box would make it true.
| Device | A$ (approx.) | Idle W | Load W | 8B tok/s | Panels | Battery | Note |
|---|---|---|---|---|---|---|---|
| Current Selma tower | ~3,000 | ~65 | ~220 | 59 | 3 × 440W | 10.2 kWh | Most efficient per token for overnight batch work |
| NVIDIA DGX Spark | 6,200–9,000 | 35–45 | 60–90 | ~20 | 2 × 440W | 5.1 kWh | High idle floor hurts solar |
| Ryzen AI Max+ 395 mini PC | ~3,300 | 12.5 | ~112 | 37 | 2 × 440W | 5.1 kWh | Easiest port of today's Ubuntu stack |
| Mac mini M5 Pro | ~3,000–4,400 Est. | ~5–8 | ~50–60 | ~45–55 Est. | 1 × 440W | 2.6 kWh | Rated to only 35°C ambient; macOS build |
| Jetson AGX Orin 64GB | ~5,300 | ~10–15 | ~50 | 20–40 | 1 × 440W | 2.6 kWh | DC-native; industrial variant for heat and humidity |
| Jetson Orin NX 16GB | ~1,500 module | ~5 | ~20–25 | ~18–20 | 1 × 440W | 1.3 kWh | 8B models only |
| Raspberry Pi 5 + AI HAT+ 2 | ~450 | 3.4 | 5.2 | n/a | 1 small | 1.3 kWh | Can't run 8B models. A search helper at best. |
NVIDIA raised Jetson prices by 57–101% in July 2026, and memory costs pushed most AI hardware up this year. Re-quote at the time of order.
The catch for Selma specifically is reading speed. Selma is retrieval-first, so each question carries 2,000–8,000 tokens of documents. The tower reads 4,000 tokens in under two seconds. A low-power box may take 5–15 seconds before the first word. Smaller chunks, better re-ranking and caching the Constitution help.
Today's tower, or bigger, in a capital city on grid power with a battery backup. Heavy batch work, many users, spares for the spokes.
Jetson AGX Orin 64GB (industrial for harsh sites) in a district office, hospital or larger school. Runs 14B models comfortably.
Jetson Orin NX 16GB in a fanless box for a clinic, village school or outreach. 8B models, modest document index.
A power-capped Ryzen AI Max+ mini PC. Faster, keeps today's software, needs twice the solar.
Hardware and solar, supply only. Add 25–50% for freight, duties and installation at a remote site. Lithium batteries ship as dangerous goods. Estimates
Scheduling batch jobs for 10am to 2pm, when the panels are producing, cuts the tower's battery need by about a third. A "solar-aware" scheduler that reads the charge controller would be a small, visible piece of engineering.
| Threat | Response |
|---|---|
| Heat (30–34°C common, 40°C+ under iron roofs) | Coolest shaded room, power-capped GPU, low-watt hardware, industrial modules rated to 85°C. A 250W tower cannot live in a sealed case. |
| Humidity and condensation | Conformal coating, silica gel with indicator cards, keep units powered, acclimatise before unboxing. |
| Salt air | Sealed or filtered enclosures, stainless mounts, marine cable, salt-mist-rated panels (IEC 61701). |
| Dust | Fanless designs or filtered positive pressure; quarterly cleaning. |
| Cyclones and floods | Rated mounts, raised installs, a hard case to pull the unit and store it. Outages are exactly when an offline tool matters most. |
| Lightning, dirty power | Surge protection on every line; the battery doubles as a UPS. |
| Theft | Locked cabinets, full-disk encryption so a stolen unit is unreadable, and a named local custodian. Community ownership is the best deterrent. |
| Spare parts take weeks | A pre-imaged spare drive at each site, one spare unit per 5–10 spokes at the hub, a trained local technician in each country. |
Solar, offline digital library in a rugged case; 600+ units, deployed in Vanuatu, Samoa, Tonga and Micronesia with Peace Corps. Its new offline Llama health chatbot (EDgeAI) runs on the same Raspberry Pi, though its pilots so far are in Rwanda and Lesotho. The clearest precedent and a natural partner.
Offline education server with AI-powered search. Search, not a cited assistant.
Open-source offline libraries and learning, used in dozens of countries. Could share a local network with Selma.
A review of 38 AI tools for community health workers found them dominated by connected chatbots. No other governed, cited, offline assistant for the Pacific or Timor-Leste was found. Worth checking before claiming "first".
| Language | Qwen3 | Llama 3/4 | Aya 101 | SEA-LION / Sailor2 | FLORES-200 (translation, non-commercial) |
|---|---|---|---|---|---|
| Tetum (Timor-Leste) | No | No | No | No | No |
| Portuguese, Indonesian (Timor-Leste) | Yes | Yes | Yes | Indonesian yes | Yes |
| Tok Pisin (PNG) | Yes | No | No | No | Yes |
| Bislama (Vanuatu) | No | No | No | No | No |
| Fijian | No | No | No | No | Yes |
| Samoan | No | No | Yes | No | Yes |
| Tongan | No | No | No | No | No |
| Khmer, Tagalog | Yes | Tagalog (Llama 4) | Yes | Yes | Yes |
"No" means no official support claimed; the model may have seen some text. Official model cards and blogs. PNG alone has ~840 living languages.
Retrieval helps: Selma needs to read and summarise local documents more than write fluently in the language. Timor-Leste can start in Portuguese, Indonesian and English. One idea from the research: build a small evaluation set per language with local speakers (say 200 questions over real local documents), published under community-controlled terms, in the spirit of Te Hiku Media. It's cheap, measurable for DFAT, and a contribution in its own right.