Qwen 3.8 Max
Qwen3.8-Max is Alibaba's largest model, released on the international API on Aug 2, 2026, with 2.4 trillion parameters, 95 billion active per token, and it is the first in the Max line with open weights. On the API it costs US$ 2 per million input tokens and US$ 6 per million output in the Singapore region, with a 1 million token window. In Alibaba's own table, it trails Claude Fable 5 on most agent tests and leads the image, document and video ones.
Qwen 3.8 Max in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Alibaba (Qwen team)
- API identifier
qwen3.8-max; pinned versionqwen3.8-max-0902, from Sep 2, 2026- Release
- Aug 2, 2026 on the international API (post dated Aug 3, 2026, Beijing time)
- Weights
- open on Hugging Face (
Qwen3.8-2.4T-A95B), under the Qwen3.8-Max License - Size
- 2.4 trillion parameters, 95 billion active (10 of 512 experts per token, plus 1 shared)
- Input
- US$ 2 per million tokens in Singapore; US$ 1.65 in Beijing, Hong Kong, Frankfurt, Tokyo and Virginia
- Output (includes reasoning)
- US$ 6 per million in Singapore; US$ 4.951 in the other regions
- Cache read
- US$ 0.25 (implicit cache) or US$ 0.17 (explicit), per million, on QwenCloud
- Explicit cache write
- US$ 2.50 per million
- Batch
- half price, but for this model only in the Beijing region
- Context window
- 1 million tokens, single price
- Max output
- 131 thousand tokens, with up to 262 thousand of reasoning
- Reasoning
- on by default and can be turned off on the API; effort low, medium or xhigh (default xhigh)
- Input and output
- text, image and video in; text out (the open weights accept text only)
- Reliable knowledge
- date not published by Alibaba
- Where it runs
- QwenCloud and Alibaba Cloud Model Studio, Qwen Studio and, with the weights, on vLLM, SGLang or TokenSpeed
Where Qwen 3.8 Max does well, where it falls behind and who it is for
Where it does well
- Image, document and video: in Alibaba's multimodal table it has the top score on 36 of 55 tests, such as OmniDocBench 1.5 (92.1) and VideoMMMU (88.7).
- Professional work: leads PRBench-Finance (58.3), PLawBench (73.2) and HealthBench (60.2) in the maker's text table.
- Following instructions and searching: 82.8 on IFBench and 81.9 on WideSearch, the top scores in those rows of the same table.
- Cheap output for its size: US$ 6 per million tokens, 40% less than the US$ 10 of Claude Sonnet 5.5 and GPT-6.1 Sol, with the same US$ 2 input.
- Open weights: it is the first time Alibaba has released a Max-line model. You can run it on your own servers and fine-tune it on company data.
Where it falls behind
- Behind the closed leaders in coding: 56.6 on DeepSWE 1.1 against 73.0 for GPT-5.6 Sol, and 67.7 on SWE-bench Pro against 80.0 for Claude Fable 5, in Alibaba's own figures.
- Loses to its open neighbor: in the table Z.ai published, GLM-5.3 scores higher on all 12 tests where both appear, and costs US$ 1.40 and US$ 4.40.
- The weights are not the API product: the downloadable version accepts text only and always reasons; images and non-reasoning mode stay on the API.
- Heavy to run in-house: about 4.89 TB in BF16 and 2.50 TB in FP8, with no published hardware recommendation.
- Tools and hard knowledge: 72.5 on Toolathlon Verified and 56.2 on HLE with tools, below Opus 4.8, Fable 5 and GPT-5.6 Sol in Alibaba's table.
Who it makes sense for
- Reading long documents with tables and charts, such as financial reports of more than 200 pages, and analyzing long videos: this is where Alibaba's table favors the model most.
- Multi-step office agents handling contracts, spreadsheets and reports, through the OpenAI and Anthropic compatible API.
- Apps that generate a lot of text, such as reports and summaries at volume, where output at US$ 6 weighs less than on closed models of the same tier.
- Companies that want a large model without depending on an API and own, or rent, infrastructure for a few terabytes of weights.
When to pick another model
- Long coding agents where every point counts: Fable 5 and GPT-5.6 Sol score higher on DeepSWE and SWE-bench Pro, even in Alibaba's table.
- High volume of simple tasks: Qwen3.8-Flash, from the same maker, costs US$ 0.15 and US$ 0.47 per million.
- Requirements that inference stay in Europe or the United States: in those regions the model only runs in the Global scope.
- Batch processing outside China: the 50% discount for Qwen3.8-Max only applies in the Beijing region.
What Alibaba published at launch
Qwen3.8-Max reached Alibaba Cloud's international API on Aug 2, 2026. The biggest news is not the score: it is the first time a Max-line model gets open weights, with the first file upload to Hugging Face on Aug 9, 2026. It is a mixture-of-experts model, with 10 of 512 experts working on each token, plus 1 shared, built on the Qwen3.5 architecture.
The text table in the announcement has 31 tests against Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol and Qwen3.7-Max. Qwen3.8-Max has the top score on 7 of them, one tied; Fable 5 on 15. Two caveats: the rivals already have newer versions, released between Sep 1 and Sep 29, 2026, and several tests are Alibaba's own, such as QwenSWEBench and CoWorkBench.
| Qwen3.8-Max Alibaba | Fable 5 Anthropic | GPT-5.6 Sol OpenAI | Opus 4.8 Anthropic | Qwen3.7-Max Alibaba | |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 terminal |
86.6 | 84.6 | 88.8 | 84.6 | 74.5 |
| SWE-bench Pro coding |
67.7 | 80.0 | 64.6 | 69.2 | 60.6 |
| DeepSWE 1.1 long-horizon coding |
56.6 | 70.0 | 73.0 | 59.0 | 21.6 |
| FrontierSWE coding |
73.5 | 88.8 | not published | 70.0 | 40.7 |
| PaperBench reproducing a research paper in code |
93.0 | 88.8 | 90.5 | 80.3 | 64.8 |
| JobBench professional work |
53.4 | 57.4 | 45.4 | 48.4 | 31.3 |
| Toolathlon Verified tool use |
72.5 | 77.9 | 74.9 | 76.2 | 49.7 |
| WideSearch broad web research |
81.9 | 81.2 | not published | 72.9 | 75.2 |
| HLE with tools hard knowledge |
56.2 | 64.5 | 58.0 | 57.9 | 53.5 |
| IFBench instruction following |
82.8 | 63.5 | 72.7 | 62.2 | 79.1 |
| PRBench-Finance finance |
58.3 | 55.8 | 55.5 | 51.9 | 46.8 |
| GPQA Diamond science |
92.6 | 92.6 | 94.1 | 92.0 | 92.4 |
Figures from the launch post and the Hugging Face card, as Alibaba published them, without the percent sign. According to Alibaba, the Fable 5 results may have used fallbacks. "Not published" cells are empty in the original table. The dot marks the best in each row.
Open weights: what changes for a company
With the weights, the model runs on your own or rented servers. Prompts never leave your network, the model does not disappear if the API changes its price, and fine-tuning on company data stays with you. Alibaba points to vLLM, SGLang and TokenSpeed for serving.
But the open version is not the same as the API one: it accepts text only, always reasons and has a native context of 262,144 tokens, extendable to 1,010,000. Images, video, non-reasoning mode and built-in tools stay on the API. And size matters: about 4.89 TB in BF16 and 2.50 TB in FP8.
The Qwen3.8-Max License allows commercial use and sets two conditions that only reach large operations:
- You may use, copy, modify, fine-tune, host and sell it for free, keeping the copyright notice, with no warranty of any kind.
- Commercial products with more than 100 million monthly active users, or more than US$ 20 million in monthly revenue, must prominently display the model name.
- Anyone selling the model as a service, or selling an AI assistant for coding or office work, with revenue above US$ 50 million over 12 months needs a separate license from Qwen.
- Internal use, which does not hand the model, its outputs or its capabilities to third parties, falls outside the second condition.
Data and regions on the official API
The Model Studio documentation says Alibaba Cloud never uses customer data to train models. The region sets where data is stored; the deployment scope, where inference runs. Singapore uses the International scope and Beijing the Chinese mainland one. In Frankfurt, Tokyo, Hong Kong and Virginia, Qwen3.8-Max is only in the Global scope, with no geographic boundary, and Alibaba itself recommends a bounded scope for anyone with compliance requirements.
Pricing: Qwen3.8-Max on the API against its neighbors
Through the API, Qwen3.8-Max matches Claude Sonnet 5.5 and GPT-6.1 Sol on input, US$ 2, and charges 40% less for output. GLM-5.3, also open, is 30% cheaper on input and about 27% cheaper on output. No maker has put the five in the same table; the independent scores are at the top of this page.
| Qwen3.8-Max Alibaba | GLM-5.3 Z.ai | Kimi K3 Moonshot | Sonnet 5.5 Anthropic | GPT-6.1 Sol OpenAI | |
|---|---|---|---|---|---|
| Input | US$ 2 | US$ 1.40 | US$ 3 | US$ 2 | US$ 2 |
| Output | US$ 6 | US$ 4.40 | US$ 15 | US$ 10 | US$ 10 |
| Cache read | US$ 0.25 or US$ 0.17 | US$ 0.26 | US$ 0.30 | US$ 0.20 | US$ 0.10 |
| Context window | 1 million | 1 million | 1,048,576 | 1 million | 1,050,000 |
| Long prompts | single price | single price | single price | single price | double input above 272 thousand |
| Downloadable weights | yes | yes | yes | no | no |
Standard pricing; Qwen3.8-Max is the Singapore region price, and its cache is implicit or explicit. Alibaba Cloud, Z.ai, Moonshot, Anthropic and OpenAI documentation checked on Oct 1, 2026.
How to test before adopting
Start with the API in Singapore, which gives 1 million free tokens for 90 days. Run ten real tasks on Qwen3.8-Max and on your current model, compare accuracy and cost per task, and also test medium and low effort, which are cheaper than the default xhigh. If the task is text only and the data cannot leave the building, then price your own servers for the weights.
Examples of using Qwen 3.8 Max
A financial report of more than 200 pages, with tables and charts, becomes a structured summary. Alibaba shows this use at launch, and the model scores 92.1 on OmniDocBench 1.5, the top score in its multimodal table.
Hours of meeting or customer service recordings become an index of people, events and timestamps. Alibaba says the model works with video longer than 100 hours, and it scores 88.7 on VideoMMMU in Alibaba's figures.
The weights run with vLLM or SGLang on the company network; internal use that does not offer the model to third parties falls outside the license conditions. The launch post cites a review that found 1,284 relevant clauses in under an hour; it is the maker's own case.
Where Qwen 3.8 Max stands out and where it falls behind, by the numbers
Qwen 3.8 Max has published results on 18 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Nov 5, 2025 and Oct 31, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 18 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 92.7% | #22 / 81 | #22 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 83 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 92.0% | #29 / 69 | #29 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 46.24 | #30 / 69 | #30 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 20.0% | #25 / 79 | #25 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 53.2% | #38 / 76 | #38 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 74.7% | #21 / 68 | #21 / 86 | 93.7% GPT-6 Astra |
| WebDev Arena | 1,671 | #10 / 75 | #10 / 82 | 1,820 Claude Opus 5.5 |
| SimpleQA Verified | 47.3% | #30 / 64 | #33 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 46.0% | #23 / 64 (tied with 1) | #24 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 38.0% | #16 / 62 | #16 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 58.0% | #21 / 70 (tied with 2) | #21 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 46.3% | #20 / 58 (tied with 2) | #20 / 67 | 100.0% GPT-6.1 Sol |
| Apex Agents | 63.3% | #11 / 49 | #11 / 50 | 82.4% Gemini 4 Argon |
| Gdp Pdf | 23.2% | #18 / 37 | #18 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 20.0% | #34 / 34 | #34 / 34 | 83.3% Claude Opus 5.5 |
| DeepSWE | 57.5% | #15 / 26 | #15 / 26 | 74.1% GPT-6 Astra |
| FrontierSWE | 17.8% | #18 / 20 | #18 / 20 | 65.5% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Qwen 3.8 Max with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Gemini 3.8 Flash | Google DeepMind | 157.20 | 1,551 | 75.8 | Sep 2, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 5, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
Qwen 3.8 Max or Muse Spark 1.3?
Muse Spark 1.3 is the model from another lab closest to Qwen 3.8 Max in capability. Across the 14 tests both took, this is the score:
| Test | Qwen 3.8 Max | Muse Spark 1.3 |
|---|---|---|
| Game puzzles | 38.0% | 25.0% |
| FrontierSWE | 17.8% | 25.8% |
| CritPt | 20.0% | 26.0% |
| Chess puzzles | 46.0% | 37.0% |
| Gdp Pdf | 23.2% | 27.6% |
| Lmca | 46.24 | 53.87 |
| SciCode | 53.2% | 59.7% |
| Apex Agents | 63.3% | 57.8% |
| Dtbench | 92.0% | 96.5% |
| WebDev Arena | 1,671 | 1,652 |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Qwen 3.8 Max costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
FrontierSWE
The score scale does not start at zero.
Other models from Alibaba
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Qwen3.8 Flash | Aug 26, 2026 | no ECI | n/a | n/a | API access |
| Qwen 3.8 27B | Aug 14, 2026 | 151.55 | 1,502 | 75.3 | Open weights (unrestricted) |
| Qwen 3.8 Max | Aug 2, 2026 | 156.88 | 1,536 | 78.5 | API access |
| Qwen3.7 Flash | Jul 27, 2026 | 144.64 | n/a | n/a | API access |
| Qwen3.7-Plus | Jun 2, 2026 | 147.30 | 1,510 | n/a | API access |
| Qwen3.7-Max | May 19, 2026 | 153.83 | 1,530 | 73.1 | API access |
| Qwen 3.6 Flash | Apr 27, 2026 | 143.05 | n/a | n/a | Unknown |
| Qwen3.6 27B | Apr 22, 2026 | 146.48 | n/a | 64.0 | Open weights (unrestricted) |
| Qwen 3.6 Max (Preview) | Apr 20, 2026 | 149.24 | 1,488 | n/a | Unknown |
Have you used Qwen 3.8 Max? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Qwen 3.8 Max
How much does Qwen 3.8 Max cost?
On the Alibaba Cloud API in Singapore, US$ 2 per million input tokens and US$ 6 per million output, at a single price up to 1 million tokens. In the other regions, US$ 1.65 and US$ 4.951. The weights are free, but running the model takes large infrastructure.
Is Qwen 3.8 Max open source?
The weights were released on Hugging Face as Qwen3.8-2.4T-A95B, under the Qwen3.8-Max License, which allows commercial use without paying Alibaba. Only large operations face conditions: the model name on screen above 100 million monthly users or US$ 20 million in monthly revenue, and a separate license for anyone selling the model as a service, or a coding or office assistant, above US$ 50 million over 12 months.
Is the open version the same as the API one?
No. The open weights accept text only, always reason and have a native context of 262,144 tokens. The API adds images and video, non-reasoning mode, 1 million tokens by default and built-in tools.
Is Qwen 3.8 Max better than Claude and GPT?
In Alibaba's own text table, with 31 tests, it has the top score on 7 (one tied) and Claude Fable 5 on 15. On image, document and video tests, it leads 36 of 55. Newer models, such as Claude Opus 5.5 and GPT-6.1 Sol, are not in the comparison; the independent scores are at the top of this page.
Is my data safe on the Alibaba API?
Alibaba Cloud says it never uses customer data to train models. But in Frankfurt, Tokyo, Hong Kong and Virginia, Qwen3.8-Max inference only runs in the Global scope, with no geographic boundary. For sensitive data, the way out is running the weights on your own infrastructure.
Where does Qwen 3.8 Max rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Qwen 3.8 Max ranks #20 out of 267 models, with an ECI of 156.88. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Qwen 3.8 Max better than Muse Spark 1.3?
Across the 14 tests both models took and Epoch AI compiled, Qwen 3.8 Max comes out ahead on 6, behind on 6 and tied on 2. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Qwen 3.8 Max best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Qwen 3.8 Max are: AIME 2024 and 2025 (OTIS) (#1 of 83); WebDev Arena (#10 of 75); Apex Agents (#11 of 49). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Qwen, Qwen3.8-Max launch post
- Qwen, Qwen3.8-2.4T-A95B card on Hugging Face (architecture and benchmarks)
- Qwen, Qwen3.8-Max License
- QwenCloud, Qwen3.8-Max page (cache and limits)
- Alibaba Cloud Model Studio, pricing
- Alibaba Cloud Model Studio, model releases
- Alibaba Cloud Model Studio, regions and deployment scopes
- Alibaba Cloud Model Studio, context cache
- Alibaba Cloud Model Studio, batch processing
- Alibaba Cloud Model Studio, thinking mode
- Alibaba Cloud Model Studio, security and privacy
- Z.ai, GLM-5.3 card on Hugging Face (table including Qwen3.8-Max)
- Z.ai, pricing
- Kimi API, pricing
- Anthropic, pricing (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6.1 Sol page (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking