Qwen3.8-Max Is Here: 2.4T Parameters, 1M Context, and the Weights Still Aren't Out

Qwen3.8-Max Is Here: 2.4T Parameters, 1M Context, and the Weights Still Aren't Out

$1.4728. That is what 40 live API calls across Qwen3.8-Max and its two predecessors cost us on 2026-08-04, the day after Alibaba shipped it. The surprise was not the 2.4 trillion parameters. It was this: 3.8-Max charges 35.6% more per token than Qwen3.7-Max and still came out roughly 4x cheaper per answer, because it stopped over-thinking. One more thing most launch coverage is getting wrong. It is not open source. Not today.

This post was first published on 2026-07-19, when Qwen3.8 was a preview with no published specs. It was rewritten on 2026-08-04 against the GA release and our own API testing.

Key Takeaways

• As of 2026-08-04, Qwen3.8-Max is API-only. Alibaba promised the weights "next week", meaning the week of 2026-08-10, on Hugging Face and ModelScope.
• We measured $0.001592 per short call versus $0.006428 on Qwen3.7-Max, despite a higher per-token price.
• Alibaba's five benchmark scores are vendor-reported. We found no independent evaluation published as of 2026-08-04.
• Image and video input are real and new. We verified both against the API and against 3.7-Max's refusal.

What Changed Between the Preview and GA

Qwen3.8-Max went generally available on 2026-08-03, and the release answered every open question this page listed two weeks ago. The preview era gave us a parameter count and a promise. The GA release added a confirmed 1M-token context window, text plus image plus video input, five published benchmark scores, and a per-token price.

Here is the old ledger of unknowns, resolved:

One item did not resolve. The parameter split is still contested. MarkTechPost's launch piece states plainly that the active parameter count was not disclosed by Alibaba, while SiliconANGLE, The Decoder and Coursiv all print ~95 billion active. We re-read the MarkTechPost article on 2026-08-04 and it still says not disclosed. Somebody's sourcing is wrong, and the honest thing to tell you is that reporting on this specific number conflicted on launch day.

We could not verify it either way. The OpenRouter /models response exposes no parameter-count field at all, so nothing in our testing confirms or refutes 2.4T total or 95B active.

Is Qwen3.8-Max Open Source? Not on August 4

No. As of 2026-08-04, Qwen3.8-Max is API-only. Alibaba has scheduled the weights for release "next week" on Hugging Face and ModelScope, and has announced no licence. Coverage keeps collapsing "available" into "open," and that distinction is the whole story. There are no weights to download today, whatever a headline says.

Three states are getting collapsed into one word. They are not the same thing:

We checked the hardest available evidence rather than taking anyone's word for it: a Hugging Face search for Qwen3.8 on 2026-08-04 returns no Alibaba model card. What it does return is four community uploads named Qwen3.8_4B_Distilled and variants, which are third-party distills, not the flagship. If you are scanning that page quickly, it is easy to mistake them for the real release.

One note on the licence, because precedent keeps getting reported as commitment. Qwen 3.5 and Qwen 3.6 both shipped under Apache 2.0. A restrictive community licence at 2.4T would still technically be "open weights" while changing what you are allowed to build with them. Until there is licence text, anyone telling you what you can do with these weights is guessing. That is also why Qwen3.8-Max is not in our roundup of open-source LLMs worth running in 2026: it does not qualify yet.

What We Measured: 4x Cheaper Per Call Despite a 35.6% Price Rise

We ran 40 billed calls against qwen3.8-max, qwen3.7-max and qwen3-max through OpenRouter on 2026-08-04, total spend $1.4728. Every cost below was computed from the returned usage object and cross-checked against OpenRouter's own billed figure. All of them matched. The headline finding runs opposite to the price change.

Start with the price. Live pricing from GET /models on 2026-08-04 puts 3.8-Max at $2.00 in / $6.00 out per million tokens against 3.7-Max at $1.475 / $4.425. That is a 35.6% rise on both sides, and the ratio is identical either way (2.000/1.475 = 6.000/4.425 = 1.3559). You can cross-check the current figures on the OpenRouter model page.

Now the same trivial prompt sent to all three models, three runs each, temperature: 0, streamed:

Two separate first-token columns are necessary because both reasoning models stream hidden reasoning before any answer text. On 3.8-Max a token arrives in under a second on two of three runs, but the first word a user can actually read lands four to five seconds later. On 3.7-Max run 1 it never landed at all. If you build a streaming UI against a reasoning model, the spinner keeps spinning long after the connection proves alive.

Read the reasoning column twice. On a prompt asking for a three-sentence explanation of a Bloom filter, 3.7-Max spent between 1,174 and 1,500 reasoning tokens. 3.8-Max spent 156 to 194. That is roughly 7x less thinking for the same answer, and reasoning tokens bill at the output rate. The result: 3.8-Max is about 4x cheaper per call and 4 to 5 times faster on wall time from our machine, network round-trip included, while costing 35.6% more per token. The sticker price went up and the bill went down.

That flips as prompts grow. Modelled from live pricing rather than measured, a 30,000-token call with 500 output tokens runs $63.00 per thousand calls on 3.8-Max against $46.46 on 3.7-Max. At 100,000 input tokens it is $203.00 against $149.71. Once most of your tokens are input, the reasoning-efficiency advantage stops paying for the price rise and 3.8-Max becomes the most expensive of the three. Which model is cheapest genuinely depends on your input-to-output ratio, which is the same lesson our LLM API pricing comparison keeps landing on.

reasoning_effort Is New, and 3.7-Max Silently Ignores It

reasoning_effort appears in 3.8-Max's supported parameters and not in 3.7-Max's. On 3.8-Max it moves the needle modestly: across three runs each, minimal produced 67, 108 and 124 reasoning tokens against high at 163, 136 and 173. Medians of 108 versus 163, on the same prompt, at temperature 0.

The finding that costs money is on the older model. Sending reasoning_effort: "low" to 3.7-Max is accepted rather than rejected, and then ignored: that call still burned 1,401 reasoning tokens, billing the same $0.006689 as the identically-shaped call we logged. No error, no warning, no effect. If you are tuning cost by lowering reasoning effort, verify it is doing something on the model you are actually calling, because an unsupported parameter here fails silently rather than loudly.

The Empty-Response Trap You Should Check Before Migrating

Our first test run used max_tokens: 400 and crashed our own script. The reason turned out to be worth more than the test. Qwen3.7-Max can spend its entire token budget on reasoning and return an empty string, with finish_reason: "length", billed in full.

We hit it three separate times. At max_tokens: 400: 402 completion tokens, 400 of them reasoning, zero answer characters, $0.001822 billed. At max_tokens: 1500: 1,502 tokens, 1,500 reasoning, zero characters, $0.006689 for nothing. And once more on a later call, $0.006735. No error, no exception, just a fluent-looking response object with an empty content string.

If you are migrating an existing 3.7-Max integration and your code sets a modest max_tokens, budget time to test this path. 3.8-Max's much shorter reasoning makes it meaningfully less exposed. It is not immune.

A Small Token Overhead Nobody Mentions

The identical 12-word prompt counted 67 prompt tokens on 3.8-Max and 29 on 3.7-Max, consistently across every run (27 on qwen3-max). That is roughly 38 tokens of fixed overhead, presumably a larger system or chat template. On a 100,000-token RAG call it rounds to nothing. On a high-volume classifier firing short prompts, it more than doubles your input bill before you have written a word.

Does Qwen3.8-Max Really Accept Images and Video?

Yes, and multimodal input is genuinely new in this generation, not a relabelling. The API reports text+image+video->text for 3.8-Max and text-only for 3.7-Max. We verified the difference by sending the same image to both, and the older model rejected it at the routing layer.

We generated the test image locally with a hand-written PNG encoder so the ground truth was known by construction: 750x270 pixels, black block digits 4739 on white, with a red horizontal bar across the top. Sent base64-encoded, qwen3.8-max returned DIGITS: 4739 and TOPBAR: red. Correct on both counts, 6.99s, $0.002146. The identical request to qwen3.7-max came back as HTTP 404 {"error":{"message":"No endpoints found that support image input"}}.

Video works too, with a caveat we nearly reported as a failure. Two of the three public MP4 URLs we tried returned HTTP 400 "Failed to download multimodal content". That is Alibaba failing to fetch the file, not the route refusing video. With a URL it could fetch, 3.8-Max described a Big Buck Bunny clip accurately, including naming the character, in 24.24s for $0.006528.

Two practical notes before you build on this. The video billed as 696 ordinary prompt tokens for a roughly 10-second clip, and usage.prompt_tokens_details.video_tokens reported 0, so you cannot break out video cost from that field. And a malformed content part fails silently: sending {"type": "video", "video": [url]} instead of video_url returned HTTP 200 with the model politely saying it could not see any video, plus a bill. Use video_url.

Worth knowing: when we asked 3.8-Max to describe its own capabilities, it replied Input modalities: text. That is wrong, as the next test in our own run demonstrated. A model's self-description is a language-model output, not a spec sheet.

How Well Does the 1M-Token Context Window Hold Up?

We planted three unique facts at 10%, 50% and 90% depth in generated filler and asked for all three in one call. 18 of 18 needles came back correct across six runs on two models, at prompt sizes from 5,723 up to 247,911 tokens, graded by exact substring match in code rather than by reading the answers.

Our qwen3.8-max runs (the two qwen3.7-max runs at 83,380 and 247,873 tokens, and one qwen3-max run at 78,255, also returned every needle):

Latency scales sub-linearly, which is the practically useful part: 43x more input tokens cost only 4.2x more wall time.

Now the honest limits, because this is the easy end of long-context evaluation. Retrieving a verbatim planted fact says very little about reasoning across a large document, and we tested each size once. We did not run a 1M-token call. At $2.00 per million input tokens, one would have cost about $2.00, more than this entire test run, and a single sample would not have told us much. The advertised 1M ceiling is therefore unverified by us. And 3.7-Max matched 3.8-Max exactly at both sizes we compared, so long-context retrieval is not where this generation separates.

One pricing trap surfaced here that the launch coverage misses entirely. qwen3-max carries tiered pricing overrides that double its rate above 32,000 prompt tokens and raise it again above 128,000, while 3.8-Max and 3.7-Max are flat-rate at every length. We confirmed the tier from actual billing: our 78,255-token call to qwen3-max was charged $0.12227, the $1.56/M rate, not the $0.78/M headline. At that length it costs almost exactly what 3.7-Max costs ($0.12523). Its headline price is less than half. Its real price at 80k tokens is a rounding error away.

Alibaba's Five Benchmark Scores, and Why None of Them Are Verified

Alibaba published five benchmark scores with the GA release. Every one of them comes from Alibaba's own internal runs, and we found no independent evaluation published as of 2026-08-04. That sentence deserves to sit next to the numbers rather than three paragraphs below them.

Alibaba also reported an internal aggregate of 0.725, up from 0.474, and positioned the model as near or above Claude Opus 4.8, Fable 5 and GPT-5.6 Sol. Arena placements circulated too: 5th in Text Arena and 2nd in Vision Arena per Alibaba's own 2026-08-03 announcement, which we have not reconfirmed on the live leaderboard. Arena ranks move daily. Treat any rank you read this week as a snapshot with a date on it.

What nobody has measured yet, including us: throughput under concurrency, non-English performance, safety and alignment evaluations, and tool-calling reliability. Our own numbers cover latency, cost, modality and long-context retrieval on one route, and nothing else. Bloomberg's launch report repeated the benchmark claims without independent testing, which is where essentially all coverage sits right now.

For numbers that have been checked, our Qwen vs DeepSeek vs GLM comparison is the better reference, and Kimi K3 at roughly 2.8T is the size rival everyone benchmarks this against.

Qwen3.8 vs Qwen3-8B, and the Open-Weight Reversal Behind This Release

Qwen3.8 is Alibaba's 2.4-trillion-parameter flagship. Qwen3-8B is a dense 8-billion-parameter model from the Qwen3 series, downloadable since 2025. Similar-looking names, roughly 300x apart in size. Search engines still conflate them, and so do a surprising number of forum answers.

The naming problem got worse this week, not better. The only things on Hugging Face carrying a "Qwen3.8" name right now are community 4B distills. If you go looking for weights and grab one of those, you are downloading something with no relationship to the 2.4T model.

Two Closed Flagships, Then This

The open-weight promise is the actual news here, and it reads differently once you see the family history. Alibaba spent two generations running a deliberate split: open-weight workhorses for everyone, closed flagship at the top.

Alibaba closed the flagship tier for two straight generations and is now saying it will open the biggest model it has ever built. If the weights land as promised, that is a real reversal and it puts pressure on every lab treating "frontier tier stays closed" as settled. If they slip, this becomes the third closed Max release in a row. Both outcomes are live on 2026-08-04. We saw the same dynamic play out with GLM 5.2, where the shipped licence mattered more than the announcement.

Can You Run Qwen3.8-Max Yourself? (Still No)

No, and the GA specs make that clearer rather than less clear. At 2.4 trillion total parameters, this is not a model you serve on your own hardware, even after the weights drop. DeepSeek-V3.2 at 685B is already described by people who have done it as a serious infrastructure project. This is several times that.

So what does an open-weight 2.4T model actually buy you?

• Inference providers can host it, which means price competition, which means your cost per token falls
• Sovereign, regulated and air-gapped deployments become possible at this tier for the first time
• Researchers and fine-tuners get a frontier-scale base model to work from
• It does not mean it runs on your machine, your single A100, or your startup's GPU budget

If you want the arithmetic on what running large open-weight models actually demands, our LLM VRAM requirements guide does that math properly. The short version for this model: don't.

Should You Switch, Test, or Wait?

The method matters more than the verdict. Every model we have production-tested has behaved differently from its leaderboard position, because your prompts, your codebase and your tolerance for retries are not in anyone's benchmark. Two coding tasks in our run make the point: 3.8-Max and 3.7-Max both scored 11 of 11 on a string-width truncation task and both passed 5,000 of 5,000 fuzzed cases on date arithmetic. On correctness we could not tell them apart. The gap we measured lives in latency and token spend on easy prompts, not in capability on hard ones.

Weighing a migration and want a second pair of eyes on the cost model? Have our team pressure-test it on your workload.

All figures verified on 2026-08-04. Pricing, arena rankings and the weights timeline change frequently. Our measurements come from a single route (OpenRouter to Alibaba), one machine, one day, with n=3 for latency and n=1 for most functional probes.

Frequently Asked Questions

Is Qwen3.8-Max open source?

Not on 2026-08-04. It is API-only. Alibaba has scheduled the weights for "next week" on Hugging Face and ModelScope, and has announced no licence. Headlines calling it open source are ahead of the facts. A Hugging Face search returns only third-party 4B distills, not an Alibaba model card.

How much does Qwen3.8-Max cost?

$2.00 per million input tokens and $6.00 per million output, with cached input reported at $0.25. That is 35.6% more than Qwen3.7-Max on both sides. We measured a median $0.001592 per short call, roughly 4x cheaper than 3.7-Max on the same prompt, because it emits far fewer reasoning tokens.

How many parameters does Qwen3.8-Max have?

2.4 trillion total, per Alibaba's announcement and every outlet covering it. The active count is contested: SiliconANGLE, The Decoder and Coursiv report ~95 billion active, while MarkTechPost states Alibaba did not disclose it. The API exposes no parameter field, so we could not verify either figure.

What context window does Qwen3.8-Max have?

The live API reports 1,000,000 tokens of context and 131,072 max completion tokens. We tested retrieval up to 247,911 tokens with no failures. We did not run a 1M-token call, because one would have cost about $2.00 and exceeded our test budget, so the top of that window is unverified by us.

Does Qwen3.8-Max support image and video input?

Yes to both, and we verified them. It read digits and a colour correctly from a locally generated PNG, and described a video accurately when given a URL Alibaba could fetch. Qwen3.7-Max rejects images outright with a 404. Two of our three test video URLs failed at the provider's download step.

Are Qwen3.8-Max's benchmark scores independently verified?

No. Terminal-Bench 2.1 at 86.6, GPQA Diamond at 92.6, PaperBench at 93.0, OSWorld-Verified at 86.1 and DeepSWE 1.1 at 56.6 all come from Alibaba's internal runs. As of 2026-08-04 we found no third-party evaluation published for any of them.

Can I run Qwen3.8-Max locally?

Realistically no, even once the weights ship. At 2.4 trillion parameters it is several times the size of DeepSeek-V3.2, which is already a genuine infrastructure project to self-host. Expect to consume it through inference providers rather than your own GPUs unless you operate a datacenter.

Should I migrate from Qwen3.7-Max to Qwen3.8-Max?

It depends on your token mix. On short prompts we measured 3.8-Max at about a quarter of the cost and four to five times the speed. On long-input workloads it costs 35.6% more and performed identically on our retrieval test. If you need image or video input, 3.7-Max cannot do it at all.

When are the Qwen3.8-Max weights being released?

Alibaba said "next week" in its 2026-08-03 announcement, naming Hugging Face and ModelScope as the destinations. No exact date, and no licence text. A companion open-weight model, Qwen3.8-27B, is reported to ship alongside them. We plan to re-check on 2026-08-10.