A new flagship model launches and everyone wants the same two numbers: how smart is it and how fast is it. Qwen4 Max Preview shipped on November 11, 2026 with an answer to neither. Alibaba published no benchmark table and no speed figure. What the internet has instead is the first hours of hands-on testing, and the headline result is a pelican drawing that took about twenty minutes.
This post is the honest speed rundown: what the early testers measured, why the numbers are this rough on launch day, how to benchmark it yourself without fooling yourself, and what a fair speed comparison against Qwen3.8 Max will actually require. We run an LLM gateway, so throughput is the metric we watch closest. Every claim here is labeled: confirmed by Alibaba’s pages, reported by a tester, or still unknown.
The model launched with no official speed number
Start with what Alibaba actually published. The launch banner on the Qwen Cloud Model Marketplace reads “Qwen4-Max-Preview Launches on Token Plan”, and the Token Plan lists the id qwen4-max-preview with text, reasoning and visual understanding as its capabilities. There is no tokens-per-second figure anywhere in that spec, no context window, no max output and no price per request. Launch tracker Cellcog lists the same gap.
That silence is normal for a preview, but it matters for how you read the community numbers below. When the vendor prints nothing, every early figure you see is one person’s run on one day against an unknown backend, not a spec. Treat them as signals, not facts.
The pelican test: twenty minutes for one drawing
The first widely shared hands-on came from a tester in the Qwen chat app who asked the model to draw a pelican, the classic quick capability check. It got there. The catch was the wait: the full answer took about twenty minutes, at a rate the tester put at roughly twenty tokens per second. He blamed the servers rather than the model, and on a launch morning with the whole internet hitting a new preview endpoint at once, that is a reasonable read.
Two things are true at once here. Twenty minutes is unusable for interactive work, and it is probably not what the model costs at steady state. A launch-day preview behind a $3 subscription is exactly the setup that produces throttled queues: everyone wants to try the new thing, capacity is finite, and the provider slows the tap rather than buying GPUs for a fifteen-day window. The same tester’s number is also self-reported, with no shared harness, so it is a data point, not a benchmark.
The forum reviews were mixed
The pelican was not the only early test. On a Chinese forum, more pelican-style checks drew mixed reviews: some outputs impressed, others lagged or stumbled. That spread is the most honest picture of a model hours old. Early adopters hit different tasks at different times of day against backend capacity that is still being tuned, so you get a range, not a verdict.
What nobody has yet is the test that matters: a controlled head-to-head against Qwen3.8 Max, or against the Western flagships Qwen4 wants to beat, on the same prompts, the same day, the same harness. Until that exists, “is it faster than the last generation” is an open question.
Why the launch-day numbers are this rough
Four reasons, and none of them are the model being slow by design.
Launch traffic. A preview that trends on launch morning puts thousands of concurrent requests on a backend sized for a ramp, not a stampede. Queues form, tokens per second drops, and a pelican becomes a twenty-minute wait. This is the tester’s own explanation and it fits.
Preview infrastructure, not production. Previews run on capacity the provider is willing to burn for fifteen days. Alibaba has not said what hardware serves qwen4-max-preview, so you cannot assume it matches the serving stack behind the production Qwen3.8 models, which have been tuned for months.
Reasoning mode burns tokens before you see output. The model is listed as a reasoning model, which usually means a thinking phase that produces tokens you may not be shown but still wait for. A drawing task that thinks for a long time looks slow even at a healthy underlying rate. Our reasoning docs explain the toggle pattern.
Unknown output length. Twenty minutes at twenty tokens per second is roughly twenty-four thousand tokens. If the model was producing a long reasoning trace plus a detailed drawing description, the wall-clock time says as much about output length as about speed. Wall time alone never tells you tokens per second unless you also know how many tokens came out.
How to benchmark Qwen4 Max Preview yourself
If you want a number you can trust, do not trust a screenshot. Run a controlled test. The Token Plan hands you an OpenAI-compatible endpoint, so a timing harness is a few lines. The pattern below measures what actually matters: time to first token, total tokens, and tokens per second over the generated portion.
import os, time
from openai import OpenAI
client = OpenAI(
api_key=os.environ["QWEN_API_KEY"],
base_url="https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1",
)
def probe(prompt, label):
t0 = time.perf_counter()
stream = client.chat.completions.create(
model="qwen4-max-preview",
messages=[{"role": "user", "content": prompt}],
stream=True,
)
first = None
n = 0
for chunk in stream:
delta = chunk.choices[0].delta.content or ""
if delta:
if first is None:
first = time.perf_counter() - t0
n += 1
total = time.perf_counter() - t0
gen = total - (first or 0)
print(f"{label}: TTFT {first:.1f}s, chunks {n}, "
f"gen {gen:.1f}s, ~{n / gen:.1f} chunks/s")
probe("Draw a pelican.", "pelican")
probe("Refactor this function for readability: <paste code>", "refactor")Read it like this:
| Metric | What it tells you |
|---|---|
| Time to first token (TTFT) | Queue wait plus thinking lead-in. High TTFT with fast streaming means you waited in line or the model thought. |
| Total generation time | Wall clock for the whole answer. |
| Chunks per second | A proxy for tokens per second when streaming. Count output tokens server-side for an exact figure. |
Run it at three times of day, off peak if you can, and log the model id and date. A single run on launch morning will lie to you in both directions.
What a fair speed comparison needs
The comparison everyone wants, Qwen4 Max Preview against Qwen3.8 Max, is not meaningful yet, for reasons that are methodological, not mysterious.
Same tasks. Speed on a code refactor is not speed on a long-document summary, and a reasoning-heavy prompt is slower than a one-line question. Pick tasks that match how you will actually use the model.
Same day, same hour. Launch-week load is not steady-state load. A number from November 11 will not describe November 25, when the preview is scheduled to end and capacity has settled. Re-run anything you care about.
Same measurement. Wall time is not throughput. The open-weight Qwen3.8 27B is behind the seven-hundred-tokens-per-second-on-an-RTX-5090 figure circulating on X, but that is a local open model on one GPU, not the cloud Max tier. Comparing a local 27B to a cloud flagship is a category error.
Where the real benchmark numbers will come from
Right now, nowhere. We checked the aggregators on November 11 and none of them have an original Qwen4 Max result.
Artificial Analysis does not list Qwen4 Max at all; its Intelligence Index runs Claude Fable 5.2, Gemini 4 Carbon, Claude Opus 5.5 and GPT-6 Astra at the top, with the newest Qwen entry being the previous-generation Qwen3.8 Max. The circulating claim that Qwen4 Max “scored 53” traces to an unconfirmed leaked screenshot, not an AA entry. BenchLM’s Qwen4 Max comparison pages show every benchmark as “Coming soon” with a covered count of zero, explicitly stating there is no shared sourced benchmark row for the model. LMArena has no confirmed Qwen4 Max Preview entry either.
So the first credible numbers will come from one of three places: Alibaba publishing a model card, an independent lab adding it to a leaderboard, or the community converging on a shared harness. Until then, the only benchmark that matters is yours, on your prompts, on the plan you plan to buy.
Measure speed on a model you can meter
Our rankings page tracks real gateway throughput for the models we serve, and Qwen3.8 Max is live now at $0.577 / $1.73 per million tokens. Create a key and run your own probe today.
Mistakes to avoid when reading early speed reports
- Taking launch-day wall time as the model’s speed. A twenty-minute pelican on launch morning describes launch-morning load, not steady-state throughput. Re-measure later in the window.
- Confusing reasoning latency with generation speed. A thinking model can stream fast and still feel slow because it produced thousands of hidden reasoning tokens first. Measure time to first token separately from tokens per second.
- Comparing the cloud Max to the local 27B. The 700 tok/s RTX 5090 figure belongs to the open-weight Qwen3.8 27B running locally. It says nothing about
qwen4-max-previewon Alibaba’s servers. - Trusting an unconfirmed leaderboard score. The “Qwen4 Max scored 53” screenshot is not on Artificial Analysis and community members call it fake. Wait for a real entry or a published model card.
- Benchmarking on the free playground and shipping on the paid API. Different backends, different quotas. Test on the Token Plan endpoint you will actually call in production.
FAQ
How fast is Qwen4 Max Preview?
Alibaba has published no speed figure. The first hands-on report, from launch morning, was about twenty tokens per second with a roughly twenty-minute wait on a pelican drawing, which the tester blamed on launch-day server load. Treat that as one early data point, not a spec. Benchmark it yourself on the tasks you care about.
Why did the pelican test take twenty minutes?
Most likely a combination of launch-day queueing, an unknown backend, and a long reasoning-plus-output trace. Twenty tokens per second over twenty minutes is a lot of tokens; if most of them were hidden reasoning, the wall-clock time says as much about output length as about raw speed.
Is Qwen4 Max Preview slower than Qwen3.8 Max?
Unknown. There is no controlled comparison yet. The 3.8 generation has been tuned for months on production infrastructure, while the preview is hours old on unannounced hardware. A fair verdict needs the same tasks, the same harness and the same time of day, none of which the early reports provide.
Where can I see official Qwen4 Max benchmarks?
You cannot yet. Artificial Analysis has no Qwen4 Max entry, BenchLM lists every benchmark as coming soon with a count of zero, and LMArena has no confirmed preview entry. The leaked “53” screenshot is unconfirmed. Watch for a published model card or a real leaderboard row.
How do I measure tokens per second on the Token Plan?
Stream a request and time it. Measure time to first token for queue-plus-thinking latency and count streamed chunks over the generation window for a tokens-per-second proxy, as in the probe above. Run it multiple times at different hours; a single launch-day run will not represent steady state.
The bottom line
The first Qwen4 Max Preview speed report is a twenty-minute pelican, and that tells you more about launch-day load than about the model. The honest position: Alibaba published no speed figure, the aggregators have no original benchmark, and the early hands-on numbers are thin, self-reported and confounded by a stampede of new users. Run the probe above on your own prompts before you believe any number, including ours. Meanwhile Qwen3.8 Max is live and metered on One AI API Key at a published rate, so you can measure a real flagship today and switch the model id when Qwen4 lands.