Is Qwen3.8-Max really ranked number one globally?

It depends which leaderboard you mean, and this is the crux of the release. On Artificial Analysis's Agentic Index, Qwen3.8-Max scores 55.4 and ranks first in the world, edging Claude Opus 5 (55.3) by 0.1 points and ending a long run in which Claude and GPT models monopolised that board. But on Artificial Analysis's overall Intelligence Index it scores 58 and ranks 9th out of 185 models, behind the flagship models from Anthropic and OpenAI as well as Moonshot's Kimi K3. The accurate framing is 'number one on the agentic benchmark', not 'the strongest model overall'.

What are Qwen3.8-Max's parameters and specs?

It is a sparse Mixture-of-Experts model built on the Qwen 3.5 architecture with hybrid attention (standard plus linear). It has roughly 2.4 trillion total parameters with about 95B activated at inference — an activation ratio near 4% — across 512 experts, routing 10 plus one shared expert per token. Maximum context is 1M tokens (the open-weight version is natively 262,144 and extensible to about 1.01M; the hosted API defaults to 1M). Input modalities are text, image and video (2 seconds to 2 hours); output is text only. Maximum output token counts are reported inconsistently (131K in one Alibaba Cloud article, 65,536 in another), so check the official docs.

How much does Qwen3.8-Max cost?

Domestically it is ¥12 per million input tokens and ¥36 per million output tokens. Internationally it is $2 input and $6 output per million tokens, with implicit cache hits at $0.25. That is roughly 60% cheaper on input and 76% cheaper on output than Claude Opus 5, with an average cost of $1.14 per task measured by Artificial Analysis. Third-party providers such as DeepInfra list it as low as $1.65 input / $4.95 output. There are also Token Plan subscriptions at ¥39 / ¥139 / ¥499 per month. One caveat: when thinking mode is enabled, chain-of-thought tokens are billed as output.

Is Qwen3.8-Max open source?

It is the first Max-class Qwen to have open weights released (Qwen3.8-2.4T-A95B), published to Hugging Face and ModelScope in mid-August 2026. However the licence is a custom Qwen3.8-Max License rather than Apache 2.0, and it attaches commercial conditions based on metrics such as monthly active users and revenue. Read the terms before commercial use — this is a material restriction rather than a formality.

What is Qwen3.8-Max's recursive self-improvement (RSI) capability?

Alibaba disclosed at the Yunqi Conference that the model can independently build training pipelines, construct training data, design experiments and locate defects, and that it iterated for over a month with zero human involvement, completing 33 effective iterations that raised its Artificial Analysis score by 12.5%. Alibaba's Qwen lead describes scaling as the path toward ASI, with RSI already applied to training, inference and chip-model co-design. Treat these figures as vendor-reported: the specific evaluation conditions have not been disclosed, so this is a directional demonstration rather than an established result.

What are the main criticisms of the 'number one agentic' claim?

Several are worth weighing. The margin over Claude Opus 5 is only 0.1 points, which sits inside evaluation noise and does not represent a stable advantage. Overseas coverage notes that while Qwen3.8-Max genuinely leads on computer-operation benchmarks like OSWorld-Verified, when all agentic tasks are considered together Claude Opus 5 and OpenAI's flagships retain an overall edge — its real standout is the performance-to-cost ratio rather than absolute capability. Some demos and benchmark results are vendor-published and have not been widely reproduced by independent evaluators. Output is text only, with no image or video generation. And the custom licence restricts commercial use.

Qwen3.8-Max Review 2026: #1 on AA's Agentic Index at 55.4 — and Only 9th Overall, Which Is the Real Story

September 23, 2026 · Updated 2026 · Agentic AI · MoE · Open-Weight LLM · Read time: ~12 min

Verdict first: Qwen3.8-Max is two true statements that look like a contradiction. On Artificial Analysis's Agentic Index it scores 55.4 and ranks first in the world — edging Claude Opus 5 by 0.1 points and ending a long stretch in which Claude and GPT models owned that leaderboard. On AA's overall Intelligence Index it scores 58 and ranks 9th of 185 models, behind the Anthropic and OpenAI flagships and Moonshot's Kimi K3. Alibaba disclosed the agentic result at the Yunqi Conference on September 22, 2026, alongside a genuine second headline: it leads OSWorld-Verified at 86.1 for autonomous computer operation, ahead of Claude Fable 5 (85.0) and GPT-5.6 Sol Max (83.2). The model itself is a sparse MoE with 2.4 trillion total parameters activating only ~95B, a 1M-token context and video input. The durable advantage, though, is the price: $2 input / $6 output per million tokens — roughly 60% cheaper on input and 76% on output than Claude Opus 5, at $1.14 per task. A single point to hold onto: a 0.1-point lead is a photo finish, not dominance.

Qwen3.8-Max review 2026 — Alibaba's 2.4 trillion parameter MoE that tops the Agentic Index at 55.4 but ranks 9th overall

Qwen3.8-Max — 2.4T total / 95B active, 1M context, video input: #1 on the agentic benchmark, 9th on overall intelligence, and 60–76% below Opus 5 on price

What Is Qwen3.8-Max?

Qwen3.8-Max is Alibaba's flagship foundation model, live on Alibaba Cloud Model Studio from August 2, 2026 and officially announced August 3. It is built on the Qwen 3.5 architecture.

The reason it deserves its own piece now is the September 22 Yunqi Conference in Hangzhou (themed "intelligence put to use", anchored on agentic AI), where Alibaba laid out the model's evaluation results and roadmap. What emerged is a very typical — and very easily misread — kind of "number one".

Two facts sit side by side and both are correct:

It is the best model in the world at agentic computer operation by the standard benchmark
It is not the strongest model overall, and by a fair margin

Everything else in this review follows from holding those two ideas at once.

Architecture and Specs

Spec	Qwen3.8-Max
Architecture	Sparse MoE, built on Qwen 3.5, hybrid attention (standard + linear)
Total parameters	~2.4 trillion
Active parameters	~95B — activation ratio ~4%
Experts	512 experts, routing 10 + 1 shared per token
Context	1M tokens max (open weights native 262,144, extensible to ~1.01M; hosted API defaults to 1M)
Max output	Reported inconsistently — 131K in one source, 65,536 in another
Input modalities	Text, image, video (2 seconds to 2 hours)
Output modality	Text only
Licence	Custom Qwen3.8-Max License (not Apache 2.0)

Read "2.4 trillion parameters" and "95B active" together. The 4% activation ratio is the entire reason this model can be priced at $2/$6 — total parameters set knowledge capacity, while activated parameters set the compute bill on every token. It is the same structural bet StepFun made with Step 5 Preview (600B/27B) and Xiaomi made with MiMo-V2.6 (1.02T/42B); the labs are converging on it.

Two Leaderboards, One Contradiction

Leaderboard A — AA Agentic Index: #1 in the world

Rank	Model	Score
🥇 1	Qwen3.8-Max	55.4
2	Claude Opus 5 (max)	55.3
3	GPT-5.6 Sol	54.0
4	Claude Fable 5	52.8
5	Kimi K3	50.1

Alibaba dates this board to August 6, 2026. Before this, the category had been locked down by Claude and GPT — the previous best from a Chinese model was Kimi K3 at 50.1.

Leaderboard B — AA overall Intelligence Index: 9th

Metric	Value
Overall score	58
Rank	9th of 185 models evaluated
Context	Comparable to Claude Opus 4.8; ahead of every Google, Meta and xAI model — but behind Anthropic's and OpenAI's flagships, and behind Kimi K3

The honest reading: Qwen3.8-Max is a specialist that wins its discipline, not an all-round champion. First on the agentic index, first on OSWorld — ninth on general intelligence. Overseas coverage makes the same point from the other direction: it genuinely leads on computer-operation benchmarks, but when agentic tasks are considered as a whole, Claude Opus 5 and OpenAI's flagships retain the overall edge. Its standout quality is the performance-to-cost ratio, not absolute capability.

And the margin: 55.4 versus 55.3. That is one tenth of a point — inside the noise floor of any evaluation. Report it as a photo finish between peers, not as a gap. Some demos and benchmark figures in this release are vendor-published and have not been widely reproduced by independent evaluators.

OSWorld 86.1: The Benchmark With Teeth

Model	OSWorld-Verified
Qwen3.8-Max	86.1 🥇
Claude Fable 5	85.0
GPT-5.6 Sol Max	83.2
Gemini 3.1 Pro	76.2

If you want a reason to take the agentic claim seriously rather than dismiss it as leaderboard-shopping, this is it. OSWorld-Verified measures an agent actually driving a computer — real desktop and web tasks, not multiple-choice questions. A 86.1 leading Fable 5 at 85.0 and GPT-5.6 Sol Max at 83.2 is a concrete, task-level result in the capability that agent products are actually built on.

Qwen3.8-Max also leads on PaperBench (93.0), TerminalBench 2.1 (86.6) and Vision2Web (69.0), and places first in the CodeArena front-end programming evaluation. The pattern across all of them is consistent: this is a model tuned for doing work, not for general knowledge recall — which is exactly what the 9th-place overall ranking reflects from the other side.

Pricing: The Durable Advantage

Item	Domestic (CNY / 1M)	International (USD / 1M)
Input	¥12	$2
Output	¥36	$6
Implicit cache hit	¥1.5	$0.25
Explicit cache creation	¥15	—
Explicit cache hit	¥1	—

Versus Claude Opus 5: ~60% cheaper on input, ~76% cheaper on output.
$1.14 average cost per task (Artificial Analysis measurement).
Third-party providers go lower still — DeepInfra and similar list roughly $1.65 input / $4.95 output per million tokens.
StepFun's competing claim lands in the same band: Step 5 Preview is $1/$2.70 with a 95% cache discount and ~$0.71 per AA task. The sub-$2 frontier tier is getting crowded, which is good news for buyers.
Watch this one: with thinking mode enabled, chain-of-thought tokens are billed as output. That means real-world cost will run above the listed rate on reasoning-heavy workloads.

There are also Token Plan subscriptions at ¥39 / ¥139 / ¥499 per month (introductory pricing, with 2,500 / 10,000 / 40,000 credits refreshed every 7 days).

Recursive Self-Improvement and the Qwen4 Roadmap

The other thread out of Yunqi is more speculative but more interesting: Alibaba said Qwen3.8-Max has recursive self-improvement (RSI) capability.

It can independently build training pipelines, construct training data, design experiments and locate defects
It iterated for over a month with zero human involvement
It completed 33 effective iterations, raising its Artificial Analysis score by 12.5%

Alibaba's Qwen lead frames scaling as the path toward ASI, with RSI already applied to training, inference and chip-model co-design.

Treat this as directional, not established. These figures are vendor-reported and the evaluation conditions behind the 33 iterations and the 12.5% gain have not been disclosed. It is a signal about where Alibaba says it is heading, not a reproducible capability claim.

The roadmap:

Model	Status	Planned scale
Qwen4	Training underway	Next-generation architecture
Qwen4.5 / Qwen5	Planned	5T – 10T parameters

Other items from the same keynote: Qwen3.8-Flash cuts training cost by nearly 90% through attention innovations; Qwen-Image-3.1 approaches the GPT-Image family; Qwen-Audio-3.1 reaches the international top tier across ASR, TTS and realtime; Qwen3.8-LiveTranslate cuts per-character latency to 2.3 seconds; Wan3.0 does single-pass 30-second video generation with structured input, with the next-generation video model due in November 2026.

Open Weights — With a Custom Licence

Item	Detail
Open weights	First open Max-class Qwen (Qwen3.8-2.4T-A95B), on Hugging Face and ModelScope since mid-August 2026
Licence	Custom Qwen3.8-Max License — NOT Apache 2.0, with commercial conditions tied to MAU and revenue
Ecosystem (one month)	56M+ downloads of Qwen3.8-related models, 1,900+ derivatives
Ecosystem (cumulative)	460+ Qwen models open-sourced, 3B+ downloads, 300,000+ derivatives
Named users	Reportedly Perplexity, Airbnb, Pinterest and Reuters building agents, assistants or in-house models on Qwen

The licence is a real constraint, not a formality. Alibaba has kept the weights downloadable while attaching commercial terms — a middle path between Apache 2.0 openness and a closed API. If you are evaluating this for a commercial product, read the licence text before you build, because MAU and revenue thresholds are exactly the kind of clauses that turn a free-looking model into a negotiation.

The Honest Take

Score: 8.0/10 as an agentic workhorse. 6.5/10 as a general frontier model. Both are correct, and picking which one matters is the whole decision.

What genuinely impresses:

Agentic #1 is a real result, not a technicality. 55.4 leading Claude Opus 5 ends a stretch where Claude and GPT owned that board, and the prior Chinese best was 50.1.
OSWorld-Verified 86.1 is the benchmark that matters for anyone building computer-use agents — it measures an agent actually operating a machine, and it leads Fable 5 and GPT-5.6 Sol Max.
The price is the most durable thing here. $2/$6, 60%/76% under Opus 5, $1.14 per task, and third-party routes down to $1.65/$4.95. Against a 9th-place overall ranking, that ratio is the actual product.
2.4T total with 95B active (4%) — the sparsity ratio is what makes the price possible.
1M context plus video input (2 seconds to 2 hours) gives long-horizon agents and long-video work somewhere to run.
Leads across agentic and coding benches — PaperBench 93.0, TerminalBench 2.1 86.6, Vision2Web 69.0, CodeArena front-end first.
RSI is a genuinely forward-looking disclosure — 33 iterations, a month unattended, 12.5% gain, applied to chip-model co-design. Even discounted for being self-reported, it says something about direction.
The ecosystem is enormous — 56M downloads in a month, 3B cumulative, 300,000+ derivatives.
First open Max-class weights — downloadable and deployable, which most flagship-tier models are not.

What should temper your excitement:

9th overall, not 1st. That is the fact most likely to be lost in the "global number one" headline.
A 0.1-point lead is a tie in practice. Depending on sampling and harness configuration, this ordering could flip on the next run.
Independent voices disagree with the framing. Overseas coverage explicitly notes that across all agentic tasks, Claude Opus 5 and OpenAI's flagships keep the overall edge — Qwen's strength is the cost ratio.
Some results are vendor-published and not widely reproduced.
Text output only. No image or video generation.
The custom licence restricts commercial use through MAU and revenue conditions.
Chain-of-thought is billed as output in thinking mode, so real cost exceeds the rate card on reasoning workloads.
Published specs disagree with each other — max output is given as both 131K and 65,536, and caching and batch pricing vary by source.
RSI numbers are unaudited and unexplained.
Qwen4 has only just started training — the 5T–10T plan is a roadmap, not a capability.

Who should use it: teams building agents, automation or computer-use pipelines — this is where it genuinely leads, and OSWorld 86.1 is credible evidence; budget-conscious teams wanting flagship-adjacent capability at 60–76% below Opus 5; front-end and code generation work; long-context and video-understanding workloads; and teams that need downloadable weights for private evaluation.

Who should skip it: anyone who needs the strongest overall model (it is 9th, not 1st); applications requiring image or video generation; commercial projects that cannot live with MAU and revenue clauses in the licence; procurement processes requiring independently reproduced benchmarks rather than a 0.1-point margin; and cost-sensitive workloads that lean heavily on thinking mode.

FAQ

Is Qwen3.8-Max really number one globally?

It depends on which leaderboard. On Artificial Analysis's Agentic Index it scores 55.4 and ranks first worldwide, edging Claude Opus 5 (55.3) by 0.1 points and ending Claude and GPT's lock on that board. On AA's overall Intelligence Index it scores 58 and ranks 9th of 185, behind Anthropic's and OpenAI's flagships and behind Kimi K3. The accurate framing is "number one on the agentic benchmark", not "the strongest model overall".

What are the parameters and specs?

Sparse Mixture-of-Experts on the Qwen 3.5 architecture with hybrid attention (standard plus linear). Roughly 2.4 trillion total parameters, ~95B activated at inference — about a 4% activation ratio — across 512 experts routing 10 plus one shared expert per token. Maximum context is 1M tokens; open weights are natively 262,144 and extensible to about 1.01M, while the hosted API defaults to 1M. Input covers text, image and video (2 seconds to 2 hours); output is text only. Maximum output tokens are reported inconsistently (131K in one source, 65,536 in another) — check the official documentation.

How much does it cost?

Domestically ¥12 per million input and ¥36 per million output. Internationally $2 input / $6 output per million tokens, with implicit cache hits at $0.25. That is approximately 60% cheaper on input and 76% cheaper on output than Claude Opus 5, at an average of $1.14 per task by AA's measurement. Third-party providers such as DeepInfra list it as low as $1.65 / $4.95. Token Plan subscriptions run ¥39 / ¥139 / ¥499 per month. Note that in thinking mode, chain-of-thought tokens bill as output.

Is it open source?

It is the first Max-class Qwen to have open weights released (Qwen3.8-2.4T-A95B, on Hugging Face and ModelScope since mid-August 2026). But the licence is a custom Qwen3.8-Max License, not Apache 2.0, and it attaches commercial conditions based on monthly active users and revenue. Read the terms before building on it commercially.

What is the RSI capability Alibaba announced?

Recursive self-improvement: the model can independently build training pipelines, construct data, design experiments and locate defects. Alibaba says it iterated for over a month with zero human involvement, completing 33 effective iterations that raised its Artificial Analysis score by 12.5%. RSI is also described as applied to training, inference and chip-model co-design. These are vendor-reported figures with undisclosed evaluation conditions — treat them as a directional signal.

What is the case against the "number one agentic" claim?

Four things. The 0.1-point margin over Opus 5 is inside evaluation noise and is not a stable advantage. Overseas coverage notes that while it genuinely leads computer-operation benchmarks, Claude Opus 5 and OpenAI's flagships retain the overall edge when all agentic tasks are considered together. Some demos and results are vendor-published and not widely reproduced. And the custom licence restricts commercial deployment.

When is Qwen4 coming?

Alibaba said Qwen4 is already in training on a next-generation architecture, with planned Qwen4.5 and Qwen5 scaling to 5–10 trillion parameters as the path toward ASI. No release dates were given. The next-generation video model is expected in November 2026.

How does it compare with Claude Opus 5?

Mixed, and it depends what you weight. Qwen3.8-Max leads the Agentic Index 55.4 to 55.3 and OSWorld-Verified 86.1 to Opus's tier, and it is 60% cheaper on input and 76% cheaper on output. But on the overall Intelligence Index Opus 5 ranks ahead, with Qwen at 9th of 185. Independent coverage suggests Opus 5 and OpenAI's flagships keep an overall agentic edge across the full task mix.

Sources

The Agentic Index and Intelligence Index figures are third-party measurements by Artificial Analysis. RSI iteration data, and some demos and benchmark results, are Alibaba's own and have not been widely reproduced by independent evaluators. Published specifications disagree across sources on maximum output tokens and batch pricing — verify against official documentation. The licence contains commercial conditions; read the original terms before commercial use.

搜索此博客

ring