Qwen3.8-Max: Alibaba Ships 2.4 Trillion Parameters, With Open Weights Promised

Alibaba launched Qwen3.8-Max on 3 August 2026: 2.4 trillion parameters, a million-token context, multimodal input, and API pricing of $2 per million input tokens against $6 output. It also committed to releasing the weights, along with a smaller Qwen3.8-27B checkpoint, within about a week.

That commitment is the story, and it is worth being precise about its status. As of writing, the weights are not out. Nothing is on Hugging Face, no licence has been published, and the activated-parameter count has not been disclosed. If this lands as announced it would be the first Max-class Qwen model ever released openly, which is a genuine departure for a company that kept Qwen3.6-Plus, Qwen3.7-Max and Qwen3.7-Plus closed and API-only. If it slips or arrives under restrictive terms, it is a price cut with good marketing.

Both outcomes are still live. Here is what can be verified today.

What is confirmed

Two of those rows do more work than the rest.

The pricing is aggressive against the closed frontier. GPT-5.6 Sol sits at $5 input and $30 output; Claude Fable 5 at $10 and $50. At $2 and $6, Qwen3.8-Max is undercutting the top tier by five to eight times on output while claiming, in Alibaba's own words, to be second only to Fable 5. For comparison inside the open tier, Kimi K3 lists at $3 input and $15 output.

The undisclosed active-parameter count is the more consequential gap. A sparse 2.4T model with 40 billion active parameters and one with 120 billion are completely different propositions to serve, and without that figure nobody outside Alibaba can model inference cost or size a cluster. Secondary reporting has circulated a figure of roughly 95 billion active, but that is not an official disclosure and should be treated accordingly.

The benchmarks, and who produced them

This is where care is needed, because the numbers circulating are almost entirely Alibaba's own.

From Alibaba's published table:

Read that table with three caveats attached, all of which are visible in the methodology rather than speculative.

The SWE-bench Pro figure was produced against a task set that Qwen itself refined, with baselines re-evaluated on the corrected benchmark. That is a defensible thing to do and it is also a reason the number is not directly comparable to scores published elsewhere.

The multimodal comparisons benchmark against Qwen3.7-Plus rather than Qwen3.7-Max, which flatters the generational improvement. The GPQA Diamond result illustrates the point from the other direction: 92.6 against Qwen3.7-Max's 92.4 is a two-tenths gain, which is to say no gain at all on that benchmark.

And independent replication is thin. Vals is currently the only substantive third-party suite to have published numbers, and its Terminal-Bench result diverges sharply from Qwen's own. Artificial Analysis coverage is still pending. One blind-reviewed architecture evaluation over 269 files across two real projects scored the Qwen3.8-Max preview at 80 out of 100 against Kimi K3 at 83, with the notable detail that none of Qwen's 44 tool calls failed.

The reasonable summary: Alibaba has shown enough breadth to justify calling this frontier class, and not enough standardised independent evidence to crown it best. Treat every figure above as a claim awaiting replication, which is the correct posture for any lab's launch-day table and is not a criticism specific to Alibaba.

Where the real strength looks likely to be

Setting the contested numbers aside, three things about this release stand up.

Agentic reliability. The tool-call result from the independent architecture evaluation is a better signal than most benchmark scores, because it measures whether the model can stay coherent while doing things rather than while writing prose. Forty-four calls without a failure is the kind of number that shows up in production as fewer retries.

Coding, in the band. 67.7 on SWE-bench Pro sits well behind Fable 5's 80.0, and that gap is real. But the FrontierSWE jump from Qwen3.7-Max's 40.7 to 73.5, and DeepSWE 1.1 going from 21.6 to 56.6, describe a generational improvement on agentic coding that is larger than anything else in the table. Whatever Alibaba changed, it changed the thing that matters for coding agents.

Multimodal input at frontier scale. Text, image and video in, with a million-token window, is a combination very few models offer. If your workload involves reasoning over documents that contain diagrams, screenshots or footage, this is a narrower field than the general leaderboard suggests.

The API surface, which is better than expected

Model quality gets the attention and the surrounding platform usually decides whether a migration is a week or a quarter. This one is unusually complete at launch.

The hosted API is both OpenAI-compatible and DashScope-compatible, which means an existing integration can often be pointed at it by changing a base URL and a key. That single detail removes most of the cost of running an evaluation, and it is the reason there is no good excuse for not testing it against your own prompts this week.

Beyond the basics it ships function calling, structured outputs, batch processing, prefix completion and fine-tuning, plus five tools exposed through a Responses API: a code interpreter, web search, a web extractor, and text-to-image and image-to-image search. Rate limits are quoted at 2 million tokens per minute and 15,000 requests per minute, which is generous enough that throughput is unlikely to be your constraint during evaluation.

Structured outputs and function calling being present on day one matters more than it sounds. A model that cannot reliably emit conforming JSON needs a repair step in front of every pipeline that consumes it, and repair steps are where latency and error budgets go to die. Combined with the tool-call reliability seen in the independent architecture evaluation, this is a model that appears designed for agentic use rather than for chat, which is consistent with where the benchmark gains concentrate.

The context handling deserves one note. The million-token window is not a flat million: maximum input is 991,000 tokens, dropping to 983,000 with thinking enabled, and maximum output is 131,000 in either mode. There is also a separate reasoning budget of up to 262,000 tokens. If you are designing around long-context ingestion, those are the numbers to build against rather than the round one on the marketing page.

Where to be cautious

The licence does not exist yet. This is the single largest open question, and it is being widely assumed away. Precedent cuts both ways: smaller Qwen models have shipped Apache 2.0, while every Max-class model to date has been closed. At least one prominent open-source developer has publicly flagged what they read as a geographic prohibition in draft terms covering the United States, the EU, the UK and Korea. That reading may be wrong and the terms may change before publication. Until the licence file exists, any plan that depends on permissive terms is a plan built on an assumption.

Verbosity inflates the effective price. Qwen models emit substantially more output tokens per task than peers. On the Artificial Analysis Intelligence Index, Qwen3.5-27B consumed roughly 98 million output tokens against about 56 million for MiniMax-M2.5 on the same evaluation. If 3.8 inherits that, the headline $6 output rate understates the real cost per completed task by something approaching 1.75x, which narrows the gap against models with denser reasoning. This is the same mechanism we worked through in the reasoning token piece, and it is why per-token comparisons mislead.

Serving it is a datacentre exercise. At 4-bit, 2.4 trillion parameters occupy roughly 1.2 TB for weights alone. An H200 carries 141 GB, so even eight cards leave awkward arithmetic before you have allocated any key-value cache. This is a multi-node artifact, not something an individual developer or most companies will run. The practical on-premise story is the 27B checkpoint, which we cover in the companion piece.

Prefix stability matters more than prompt length. Cached input at $0.25 costs eight times less than fresh input at $2.00. On a workload with a large fixed preamble, how you order your prompt will move the bill further than which model you pick. Put everything stable at the front in a fixed order. Retrieval systems that concatenate chunks in relevance order destroy their own cache on every request.

How to read a launch-day benchmark table

Worth building the habit, and this release is a clean case to practise on.

Treat release-day numbers as a claim rather than a result. It is not that labs lie. It is that an evaluation run by people who want a number to look good produces a different figure than an adversarial one run by people who do not care, and the second kind arrives weeks later. Alibaba's previous releases have held up reasonably under replication, drifting a point or two rather than collapsing, which earns more credibility than a first-time lab would get. Drift of a point or two is still drift.

Three questions to ask of any table like the one above. Which baseline was used, since comparing against a vendor's own weaker variant inflates the generational delta, which is exactly what happened here with Qwen3.7-Plus standing in for Qwen3.7-Max. Was the benchmark modified, since a refined task set can be entirely legitimate and still make the number incomparable to published scores elsewhere. And did the same harness run every model, since differences in agent scaffolding move agentic scores more than model quality does.

The practical move is to weight benchmarks by how far you can verify them yourself. Reasoning and maths benchmarks are hard to game and easy to reproduce, so a strong showing there is fairly trustworthy. Coding and agentic benchmarks are noisy and highly sensitive to harness details, so discount them and run the model on your own repository before believing any figure.

That is the whole argument for open weights in one sentence. When you can download the model, you can run your own evaluation on your own data and trust what you measure over what anyone published, and that option is worth more than every number in the table. Which is precisely why the unshipped weights, rather than the benchmark scores, are the thing to watch here.

What the release signals

Three things, in ascending order of importance.

The competitive pressure has moved inside the open tier. A year ago the interesting comparison was open against closed. Now it is Qwen against Kimi against DeepSeek, with each release compressing the one below it. Alibaba did not preview an open flagship out of generosity. It did so eleven days after Moonshot shipped Kimi K3 at 2.8 trillion parameters with weights on the day.

Pricing at the top is being set from below. $2 and $6 for a model claiming near-frontier capability is not a response to OpenAI or Anthropic. It is a response to what open weights have already done to the floor, a dynamic we traced in the efficiency race.

And the case for building anything durable on a base model keeps weakening. When frontier-adjacent capability arrives at a fifth of the price every few months, having it is not an advantage, because your competitors have the same option on the same day. What survives is what you build on top, which is the argument in fine-tuning on your own data.

What to do now

Evaluate the hosted API today. It is available, it is cheap, and the only way to know whether the contested benchmarks translate to your work is to run your own prompts through it. That costs very little at $2 and $6.

Do not commit to the open-weight plan yet. Wait for the weights, the licence and the activated-parameter count. All three are needed before anyone can honestly size a deployment, and two of them may not say what people are assuming.

Keep DeepSeek in the comparison. For reasoning-heavy, cost-sensitive workloads that do not need multimodal input, DeepSeek's V4 line remains cheaper on both sides of the bill.

Do not route everything anywhere. Qwen3.8-Max is strong on agentic and multimodal work and clearly behind Fable 5 on hard coding. Route the coding slice to what wins it and the volume to what is cheap, which is the multi-model position that keeps being correct. Swfte Connect runs that split behind one API, including to a Qwen endpoint you host yourself if the licence eventually permits it.

The weights are the thing to watch. Everything else in this post is a number someone wants you to believe.

---

Related: running the 27B locally, Kimi K3 self-host guide, and the live model leaderboard.