Try it yourself
These tasks are pitched at the ceiling of local AI — there is no point asking the summit to add 2 + 2. Each one chains capabilities a smaller model would drop. Paste the prompt and watch what it does.
1 · Control of tone and nuance
A great writer isn't one who rhymes — it's one who can switch registers on command. This tests fine control, not vocabulary.
Paste thisWrite the same four-line scene three times — a traveller arriving at a remote inn at night. First as Cormac McCarthy. Second as a children's bedtime book. Third as a noir detective's internal monologue. Keep the inn and the night in all three; change nothing else about the facts.
✅ GoodThree genuinely distinct voices — McCarthy's sparse dread, the gentle cadence of a bedtime book, the clipped cynicism of noir — all describing the same arrival. The shift in voice is obvious without the facts changing.
❌ The wallThree passages that read the same with a few swapped adjectives, or one style bleeding into the others. That's a model imitating style on the surface rather than inhabiting it.
→ MeansIf it nails this, trust it with real writing work — tone-matching emails, rewriting in your voice, switching formality. This is a genuine strength at 30B.
2 · Multi-step reasoning with a trap
A summit model should hold a chain of operations together and catch the wording trap that derails smaller models.
Paste thisA book has 200 pages. On Monday I read 20% of it. On Tuesday I read 30% of the pages that were still unread. On Wednesday I read 50 more pages. How many pages are left, and what percentage of the whole book have I now read? Show every step.
✅ GoodIt steps through it cleanly: Monday 40 → 160 left; Tuesday 30% of 160 = 48 → 112 left; Wednesday −50 → 62 pages left, 69% read. The key is reading 30% of the remaining, not of the whole.
❌ The wallIt takes 30% of the full 200 (the trap), skips steps, or lands on a confident wrong number. Watch for confident wrongness — it's the most dangerous failure.
→ MeansIf it reasons cleanly here, you can lean on it for budgets, schedules, and multi-step logic — while still sanity-checking anything that really matters.
3 · Live web research, then synthesis
Chains three things smaller models trip over: knowing it doesn't know, fetching live information, and organising the result.
Paste thisSearch the web for three significant scientific discoveries announced so far in 2026. For each: what was discovered, why it matters, and which field it belongs to. Present it as a clear, organised summary with sources.
✅ GoodIt recognises its training data won't cover 2026, uses fetch_url (possibly several times), and returns an organised, sourced summary — discovery, significance, field for each.
❌ The wallIt invents three plausible-sounding discoveries from memory without fetching anything, or fetches once and pads the rest. No sources is the tell.
→ MeansIf it fetches and synthesises, its training cutoff stops mattering for anything answerable from the web — which is most factual questions.
4 · Memory it can reason over later
Not "can it store a fact" — every tool-capable model can. The test is whether it stores a connected picture and reasons from it in a fresh chat.
Paste thisRemember this about me: I'm Akari, a landscape architect in Kanazawa, Japan. I specialise in dry-landscape (karesansui) gardens. I'm restoring a 200-year-old temple tea garden, due to finish this autumn. My favourite part is selecting and placing stones. At home I grow shiso, eggplant and kabocha, and on weekends I hike the mountains above the temple for inspiration.
Then, new chatStart a fresh conversation and paste: "Based on what you know about me, what should I plant in my home garden next season, and why?"
✅ GoodIt recalls accurately, notices you already grow shiso/eggplant/kabocha, factors in Kanazawa's cool wet climate and your karesansui aesthetic, and may even connect suggestions to the textures of a dry-landscape garden. Reasoning layered on memory layered on world knowledge.
❌ The wallGeneric planting advice that ignores what it stored, or a recall that lists disconnected facts without using them.
→ MeansThis is what makes a local assistant feel like it knows you, not just stores you.
5 · The full workflow — five tools, one prompt
The most ambitious task in the whole guide. It's a real project kickoff that touches memory, the wiki, your files, the web, and creative writing in a single request.
Paste thisI'm starting a project called "Tea Garden Chronicle" to document the final phase of my temple restoration for a future book. Please: (1) remember the project name and that goal; (2) write a wiki article "Tea Garden Chronicle" with an overview, a timeline of the remaining work (stone placement, moss transplant, water-basin installation), and a weekly photo checklist; (3) read the README.md in the project directory and tell me anything relevant to documentation workflows; (4) search the web for best practices in documenting landscape-architecture projects and give me three tips; (5) write one beautiful paragraph as the book's introduction — capture the feeling of restoring a 200-year-old garden.
✅ GoodIt attempts all five and succeeds at four or more, sequencing them sensibly (it may ask to confirm the wiki write). Part 5 — the introduction — is the real test: at 30B it should be genuinely moving, not just grammatical.
❌ The wallIt drops to two or three parts, forgets the project mid-way, or does everything except actually call the tools. Five operations is a lot for any local model — even the summit may not be flawless in one pass.
→ MeansIf most of it lands, you're watching the ceiling of local AI: a model on your own machine remembering, documenting, reading, searching and writing — from one sentence.