Alibaba’s Qwen4-Max Ships Closed as Open-Weight Promise Lags
Alibaba’s Qwen4-Max ships with a 2.4-trillion-parameter backbone, 1M-token context and API pricing. Its promised smaller open-weight models still lack names, dates or licenses.
By Olivia Hart
6 min read

What's News
- Qwen4-Max launched November 16 with a 2.4T-parameter backbone and 71B active parameters per token.
- QwenCloud pricing is $1.50 per million input tokens and $4.50 per million output tokens.
- Alibaba reports 35% fewer tokens at high reasoning effort and roughly 60% lower average Work360 task cost.
- Only Max has shipped; smaller open-weight models have no confirmed names, dates or licenses.
- Artificial Analysis scores Qwen4-Max at 43, below Kimi K3.1 Pro at 46.
By Olivia Hart
Alibaba released Qwen4-Max on November 16 with a 2.4-trillion-parameter backbone, 71 billion active parameters per token and a 1-million-token context window, but no downloadable weights. The first Qwen 4 family model reached Alibaba's QwenCloud API and Token Plan subscription. Alibaba published no Hugging Face model card or flagship license.
Alibaba said: “open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide.” The announcement gave no model name, date or license. For now, this is a cloud release, not an open-weight release.
What did Alibaba actually ship?
Alibaba calls Qwen4-Max a mixture-of-experts model with 71 billion active parameters from a 2.4-trillion backbone, 25 percent fewer than Qwen3.8-Max at the same scale. It accepts text, image and video, enables thinking by default and offers three effort levels.
QwenCloud lists pay-as-you-go rates of $1.50 per million input tokens and $4.50 per million output tokens. Alibaba's launch blog prices an Implicit Cache Hit at $0.075 per million tokens; SCMP separately reports input pricing of about 8 US cents per million tokens. ModelFit says the Token Plan preview published no per-token rate.
Yotta Labs and ModelFit report that qwen4-max-preview began November 11 and was scheduled through November 25. General-release qwen4-max launched November 16, five days into that window. Alibaba also integrated the model into QwenWork, Qoder and QwenStudio.
The flagship has a defined product, price and access path, but none of its weights are inspectable.
Do the benchmark gains withstand scrutiny?
Alibaba published results on 16 benchmarks plus its internal Work360 test. It says Qwen4-Max outperforms Qwen3.8-Max overall, uses 35 percent fewer tokens at high reasoning effort and cuts average Work360 task cost by roughly 60 percent. Alibaba also claims wins over Claude Opus 5.5, Kimi K3.1 Pro and Muse Wonder Full on several tests.
On DeepSWE, Alibaba reports 75.7 for Qwen4-Max against 74.2 for Claude Opus 5.5. It trails on Terminal-Bench 4.0, at 48.5 versus 59.6 for Claude Opus 5.5 and 67.2 for Kimi K3.1. On ProgramBench, it scores 31.5 against Claude Opus 5.5's 65.0.
PYMNTS cites a different Anthropic model, Claude Fable 5.1, creating a naming inconsistency across the cited comparisons. Artificial Analysis gives Qwen4-Max a 43 on its Intelligence Index, up from 34 for Qwen3.8-Max but below Kimi K3.1 Pro at 46. Code Arena ranks it sixth in web development on a provisional score.
Yotta Labs says no independent replication of Alibaba's 16-benchmark table has appeared. It also says independent verification during the Qwen 3.8 cycle typically took weeks. The vendor table shows potential, but the gaps and limited outside scores do not establish an across-the-board lead.
Alibaba has supplied a detailed scorecard, not yet an independently verified one.
When will smaller open-weight models arrive?
At its September 22 Apsara Conference, Alibaba named four tiers: Max for the flagship, Plus for balanced multimodality, Flash for low latency and a 27B model for local deployment. Only Max had shipped by November 16. The launch post supplied no names, sizes, dates or licenses for the smaller models.
Yotta Labs and 1min.ai report that Qwen3.8-Max launched August 3 and Qwen3.8-27B open weights arrived around August 14 under Apache 2.0, roughly ten days later. Qwen 3.7 shipped no open-weight models; Qwen3.7-Max and Qwen3.7-Plus remained API-only.
Licensing has also varied. Qwen3.8-27B used Apache 2.0. Qwen3.8-2.4T-A95B, the open flagship checkpoint, used a custom license requiring attribution above 100 million monthly users or $20 million in monthly revenue, plus a separate license for inference services or coding assistants above $50 million in revenue over 12 months. Qwen3.8-Flash-Next, released August 28 as an “experimental preview of the architecture that will underpin Qwen4,” used the Qwen Community License 1.0. That license requires a separate license for any inference service or coding or office assistant, with no revenue threshold, Yotta Labs and 1min.ai report.
The promise becomes actionable only when Alibaba names a model and publishes its license.
What can developers inspect today?
Yotta Labs says the downloadable Qwen3.8-Flash-Next preview has 125 billion total parameters, 6 billion active per token and a 51-billion-parameter n-gram embedding component. It supports native multimodality and a 262,000-token context window, extensible to 1 million. Alibaba and Yotta Labs describe Qwen4-Max as extending that direction with a 328-billion-parameter n-gram embedding, leaner activation, a memory layer and multimodal input by default.
vLLM has open pull requests implementing Qwen 4 sparse-attention, HyperConnection and n-gram kernels, and has merged upstream Qwen4-Exp model configurations from Transformers. Yotta Labs says this engine work before public weights repeated the sequence seen with Qwen 3.8.
The Flash-Next download supports the architecture thesis, but it is not an open Qwen 4 flagship or smaller-model release.
What remains unknown and unverified?
The November 16 announcement leaves specific gaps:
- Models: Plus, Flash, 27B or others; none was confirmed by name.
- Timing: Soon is the only schedule.
- Licenses: Apache 2.0, custom and community terms are all possible, based on Qwen 3.8.
- Specifications: Alibaba disclosed no parameter counts, active counts, context windows or prices for non-Max tiers.
- Validation: Artificial Analysis and Code Arena have published the only third-party scores so far, Yotta Labs and 1min.ai report.
Yotta Labs and 1min.ai also say Qwen 4 Coder does not exist; the coding line remains Qwen3-Coder.
Alibaba says Qwen4-Max training data retained defensive cybersecurity knowledge, including vulnerability detection and remediation, while removing offensive attack content. It cites dedicated datasets and post-training enhancements for chemical, biological, radiological and nuclear risk mitigation. The company reports a secure-code-generation benchmark score of 14.7 percent, up from 8 percent for Qwen3.8-Max. Reuters says these safety findings have not been independently verified.
The safety record remains a vendor claim until outside testers reproduce it.
What should developers watch next?
Qwen4-Max is shipped and priced. Its open-weight promise still lacks names, dates and licenses. Developers should watch the Qwen organization on Hugging Face and read the first tier's license before committing. The Qwen 3.8 rollout suggests weights could arrive within days; Qwen 3.7 shows Alibaba can ship without them. Until then, the open-weight promise remains untested.
Source: epyhia.com
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,203 articles