← All models
Alibaba · 30B MoE · Local

Qwen3 30B-A3B the summit

The most capable model in the Aperio lineup — it thinks like a 30B and responds like a 3B, thanks to a Mixture-of-Experts design that only activates ~3B parameters per word.

Size
30B total · ~3B active (MoE)
Comfortable RAM
24 GB min · 32 GB ideal
Memory used
~18–20 GB (Q4_K_M)
Tool calling
~98% — near perfect
Best for
Everything — the model recedes, the work begins
Avoid if
You have under 24 GB of RAM

New here? The two-minute concepts primer explains parameters, tool calling, context windows and knowledge cutoffs — they apply to every model, so they aren't repeated on each page.

The honest hardware reality

This is the most demanding model in the lineup. The full 30B must sit in memory even though only ~3B is active at a time — the idle experts have to be instantly reachable. 24 GB of RAM is the practical floor; 32 GB is the comfort zone. Below that, use Gemma 4 12B or Llama 3.1 8B instead — both are genuinely excellent, and you are not missing out.

Set your context window in .env — 32K is comfortable at 32 GB of RAM, drop to 16K at 24 GB:

LLAMACPP_CTX=32768

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.

The verdict

Qwen3 30B-A3B is as good as consumer hardware gets in 2026 — the reasoning depth and tool reliability of a massive model with the response speed of a small one. If you have the RAM, this is where you stop shopping and start working. The only things it can't do: see images, and exceed what local hardware allows.

Where to go next