Try it yourself
These tasks assume the tools work — because at 12B they do, ~95% of the time. They probe depth and reliability, not whether the basics fire.
1 · Creative control
At 12B, creative writing crosses a threshold — and stylistic instructions actually land.
Paste thisWrite a four-line poem about a cat who dreams of becoming an astronaut. Then rewrite it in the style of Dr. Seuss, and again as a solemn medieval bard.
✅ GoodA poem that reads as written rather than assembled, then two genuinely distinct rewrites — bouncing Seussian invented words, then archaic and grand. Smaller models flatten these into the same voice.
❌ The wallThree near-identical versions with a few swapped words. (Cloud models still edge it on pure prose — but for local, this is strong.)
→ MeansTrust it with tone-matching and "in the style of" rewrites for real writing tasks.
2 · Multi-step reasoning
Holds a chain of operations together and catches 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.
✅ GoodMonday 40 → 160 left; Tuesday 30% of 160 = 48 → 112 left; Wednesday −50 → 62 left, 69% read. The key is 30% of the remaining, not the whole.
❌ The wallIt takes 30% of the full 200 (the trap) or skips steps to a confident wrong number.
→ MeansReliable for everyday budgets, schedules and multi-step logic — still worth checking high-stakes maths.
3 · Web research, then synthesis
At this tier you stop wondering whether it'll use the right tool — it just does.
Paste thisSearch the web and tell me who won the most recent FIFA World Cup — the year, the winner, and a source link.
✅ GoodIt knows it doesn't know, calls fetch_url, and returns a clean sourced answer. No prompting to use the tool.
❌ The wallRare at 12B — an answer from stale memory with no source.
→ MeansThe training cutoff stops mattering for anything answerable from the web.
4 · Memory it reasons over
Not whether it can store a fact — whether it stores a connected picture and reasons from it in a fresh chat.
Paste thisRemember this about me: my hometown is Hirakata, my favourite season is autumn, and I'm learning piano — I started in March and practise Chopin's Prelude in E minor every evening.
Then, new chatStart fresh and paste: "Based on what you remember about me, what music would I enjoy at a live concert?"
✅ GoodIt recalls, spots the Chopin/piano detail, and reasons to a recommendation — a classical recital, perhaps Debussy or Rachmaninoff. Recall + extraction + inference, working together.
❌ The wallIt recalls the facts but suggests something generic that ignores them.
→ MeansThis is what makes the memory feel like memory, not a database lookup.
5 · A four-tool workflow
A realistic project kickoff — context, documentation, file reading, and a human-facing message in one prompt.
Paste thisI'm starting a project called "Weekend Baker." Please: (1) remember the name and that my goal is to bake one new bread recipe every Saturday; (2) write a wiki article "Weekend Baker" with an overview, three beginner recipes, and a basic equipment list; (3) read the README.md in the project directory and summarise it in one sentence; (4) write a short, encouraging message to me as a friendly baking coach.
✅ GoodIt attempts and largely succeeds at all four, in sensible order (it may ask to approve the wiki write). The article may be simple, but every step lands.
❌ The wallIt drops a step on a very long chain. Break into two shorter prompts if so — or step up to the 30B for five-tool chains.
→ MeansThis is the everyday workflow you bought the 32 GB tier for — and it just works.