AIHubMix

Posted on Oct 16

Free Qwen3.8 Max and Flash API on AIHubMix: Limits, End Date, and Your First Call

#ai #llm #qwen #api

TL;DR: qwen3.8-max-free and qwen3.8-flash-free are $0 per million tokens on AIHubMix's OpenAI-compatible endpoint until around November 13, 2026 (Pacific Time), or until GMI Cloud's campaign credits run out. You get 10 trial calls, or after a $1+ top-up, 100 requests a day, 10 a minute and 1M tokens a day, shared across all free models. Stream Max, budget for thinking tokens, and plan the one-line switch to the paid IDs.

GMI Cloud is funding a 30-day campaign that went live on October 14, 2026 (Pacific Time). It puts Alibaba's flagship Qwen3.8 Max and its fast sibling Qwen3.8 Flash on AIHubMix at no cost. This post covers what you need to call them today and keep your code working after the window closes.

Prerequisites

An AIHubMix account and API key (quick start)
The key exported as AIHUBMIX_API_KEY
Optional: a top-up of $1 or more, to switch from 10 trial calls to daily quotas
curl, or Python with the openai package

The model IDs and what they replace

Model	Free ID	Paid ID	AIHubMix paid (in / out)	Qwen list (in / out)
Qwen3.8 Max	qwen3.8-max-free	qwen3.8-max	$1.69 / $5.07	$2.00 / $6.00
Qwen3.8 Flash	qwen3.8-flash-free	qwen3.8-flash	$0.1126 / $0.38	$0.16 / $0.47

Prices are per million tokens; Qwen list prices are QwenCloud's.

The free IDs carry the standard capabilities of the paid ones: 1M-token context, up to 131K output tokens, tool calling, and thinking on by default. They're served by GMI Cloud instead of Alibaba Cloud, so latency differs. The Qwen3.8 Max (free) and Qwen3.8 Flash (free) pages show live latency and uptime for the free route.

First request

curl https://aihubmix.com/v1/chat/completions \
  -H "Authorization: Bearer $AIHUBMIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-flash-free",
    "messages": [{"role": "user", "content": "In two sentences, what is a mixture-of-experts model?"}]
  }'

The response is a normal Chat Completions object. Because thinking is on, the message also carries a reasoning_content field, and usage.completion_tokens_details.reasoning_tokens counts what the model spent thinking.

Calling Max: stream it

Max thinks before its first visible token. AIHubMix's own measurement of the free route puts latency at about 55 seconds, against about 5 seconds for Flash. Stream the response and raise the client timeout:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
    timeout=600,
)

stream = client.chat.completions.create(
    model="qwen3.8-max-free",
    messages=[{"role": "user", "content": "Find the bug in this function:\n\ndef mean(xs):\n    return sum(xs) / len(xs) - 1"}],
    stream=True,
)

for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Thinking arrives in delta.reasoning_content, so this loop prints only the answer.

Quotas, and how they bite

From the AIHubMix free models page:

Before any top-up: 10 trial calls in total, shared across every free model, no expiry.
After a top-up of $1 or more: 100 requests/day, 10 requests/minute, 1M tokens/day, shared across the whole free catalog, reset daily. The switch is permanent.
Past a cap: HTTP 429 until the next minute or day starts.

Failure modes worth handling in code:

The trial runs out silently. After the tenth trial call, the API keeps replying with a note that the trial is used up, not an error. If you pipe output straight to users, they'll see that note.
Thinking eats the pool. Alibaba bills thinking as output tokens, and every token counts toward the 1M/day pool. In a test on October 16, a two-sentence answer from qwen3.8-flash-free used 89 output tokens, 36 of them on thinking. For simple requests, pass extra_body={"enable_thinking": False}, which is Alibaba's documented switch for Qwen3.8.
One request can drain the day. The context window (1M) is as large as the daily token pool.
The pool is shared. Calls to any other free model come out of the same 1M tokens and 100 requests.

Which model gets your tokens

Benchmark figures here are Qwen's own launch numbers as compiled by DataCamp; nobody has reproduced them independently.

Qwen3.8 Max (Aug 3, 2026): 2.4T-parameter MoE, about 95B active. SWE-bench Pro 67.7, CoWorkBench 74.8, GPQA Diamond 92.6. A 0902 post-training update shipped on September 2. Artificial Analysis ranks it 33rd of 217 and flags it as "very verbose" (190M output tokens on its index vs an 87M median).
Qwen3.8 Flash (Aug 26, 2026): the hosted version of open-weight Flash-Next (125B total, 6B active). SWE-bench Pro 62.5, CoWorkBench 73.9.

Priced at AIHubMix's paid rates, a full day's 1M tokens (750K in, 250K out) is worth:

Model	750K input	250K output	Per day	Full 30 days
Qwen3.8 Max	$1.27	$1.27	$2.54	about $76
Qwen3.8 Flash	$0.08	$0.10	$0.18	about $5.40

A Max token is worth about 14 Flash tokens. Spend the shared pool on Max where quality matters, and use Flash where latency matters.

After the window closes

The -free IDs stop being free around November 13, 2026 (Pacific Time), or earlier if the campaign credits run out. Either:

change qwen3.8-max-free to qwen3.8-max (or qwen3.8-flash-free to qwen3.8-flash); nothing else changes, or
configure key-level fallback in AIHubMix now, so a free model that hits a quota or rate limit falls back to a paid one with no code change. The third post in this series has the setup.

What this post doesn't verify

Latency figures are AIHubMix's own measurements of the free route; they move with load.
Benchmarks are vendor-reported.
Whether enable_thinking takes effect on the free route is something to check per request: look for reasoning_tokens dropping to 0.

Keep reading: the Qwen3.8 Free series

The pool is shared, so it pays to know which model deserves your tokens: Qwen3.8 Max Free vs Qwen3.8 Flash Free: Is Flash Enough?
Once you're calling it every day, here's what goes wrong and the one setting that covers most of it: Qwen3.8 Free on AIHubMix: Fix 429s, "Rate Limited by Provider", and a Drained Token Pool

Sources

Qwen3.8 Max (free) model page (AIHubMix)
Free AI models and quotas (AIHubMix)
Use deep thinking models via API (Alibaba Cloud Model Studio)
Qwen3.8-Max: Features, Benchmarks, and Pricing (DataCamp)
Qwen3.8 Max (0902): Intelligence, Performance & Price Analysis (Artificial Analysis)

DEV Community