Qwen3-Next 80B-A3B review: fast answers without a reasoning pass
Qwen3-Next 80B-A3B Instruct (qwen3-next-80b-a3b-instruct) answers in seconds because it has no reasoning pass, and the quality holds up. It explained politeness corrections down to the grammar and wrote detailed summaries. Its weak spot shows up on problems with no clean answer: it thinks out loud in the response itself, and the output gets long.
What it is
Qwen3-Next 80B-A3B Instruct is an open model Qwen released in September 2025.
- Mixture of experts: 80B total, about 3B active
- Context window: about 262K tokens
- Input and output: text only
- The Instruct variant, which does not emit a reasoning trace. Tool calling and structured output are supported
Five tasks, side by side
We sent the same five Japanese-language tasks to several models.
| Task | Result | Response time |
|---|---|---|
| Apology email for a delayed delivery | Polite and natural, but it closed with 敬具 (keigu, the formal sign-off) without opening with its partner 拝啓 (haikei) | about 3 s |
| Three-point summary of an article | Detailed, with headings; it caught the currency-exchange point | about 3 s |
| Fix two politeness errors in a sentence | Both correct, and the reasons were exact: 申す (mōsu, "to say") is a humble form you use for yourself, while お〜になる (o-…-ni naru) is an honorific for others | about 4 s |
| Math word problem | Concluded there was no solution, but wrote out its trial and error along the way, about 2,650 tokens | about 12 s |
| Haiku with a seasonal word | Used 月見草 (evening primrose) and called it an autumn word; it is a summer one | about 1 s |
Most tasks came back in a few seconds. The reasoning models in the same run sometimes took 20 seconds to over a minute on these prompts.
The word problem deliberately has no whole-number answer. Qwen3-Next second-guessed itself on the page ("a calculation error?", "a mistake in the problem?") before deciding there was no solution. It did not invent an answer, which counts for a lot, but because the Instruct model has no separate reasoning channel, all of that working ends up in the reply you pay for and show to users.
Token counts for Japanese text
If you process Japanese, it is also economical with tokens. The same instruction of about 320 Japanese characters came to 218 input tokens, against 261 on GPT-4o mini and 250 on Llama 3.1 8B.
Cost
- It is billed per million input and output tokens. Current prices are on the model page.
- With no reasoning tokens, a short answer is billed as a short answer.
These runs were made in June 2026. FastMetal has since moved this model to a different upstream host; the model and context window are the same, but check the behavior on your own prompts.
Good fit, poor fit
Good fit
- Chat where response time matters
- Drafting and proofreading formal business writing
- Summarizing long documents (up to about 262K tokens)
- Agents that use tool calls
Poor fit
- Hard problems where you want a short, clean answer; a reasoning model organizes its work out of sight
- Generation that depends on fine cultural detail, such as seasonal words in haiku
FAQ
Q. Is Qwen3-Next 80B-A3B Instruct a reasoning model?
No. The Instruct variant answers directly without a reasoning trace, which is why it is fast.
Q. Can it take images?
No, it is text only.
Q. Does it support structured output?
Yes. With a JSON Schema in response_format, it returned JSON that matched the schema.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-next-80b-a3b-instruct",
"messages": [{"role": "user", "content": "Summarize the following article in three bullet points: ..."}]
}'
The same API key and balance work for every other model; change model to compare. The full list is in the model catalog.