Alibaba Launches Qwen4-Max With Closed Weights, Promises Open-Weight Smaller Models "Soon"
Alibaba released Qwen4-Max on November 16, the first model in its Qwen 4 family, making it available through its QwenCloud API and Token Plan subscription. The flagship model is closed-weight: Alibaba published no downloadable weights, no model card on Hugging Face, and no license. In the same announcement, the company said "open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide" — without naming a model, a date, or a license [8].
What shipped
Qwen4-Max is a mixture-of-experts model with a 2.4-trillion-parameter backbone and 71 billion parameters active per token — 25 percent fewer active parameters than Qwen3.8-Max at the same total scale [1][4][8]. It supports a 1-million-token context window, native text, image, and video input, and has thinking enabled by default with three effort levels.
Alibaba's QwenCloud lists a pay-as-you-go price of $1.50 per million input tokens and $4.50 per million output tokens [1][3]. Separate cached-input pricing applies: the company's own blog notes an "Implicit Cache Hit" rate of $0.075 per million tokens [8], and SCMP reports input pricing of about 8 US cents per million tokens [4]. The preview was Token Plan only with no per-token rate published [2].
Access runs through Alibaba's Token Plan subscription and the QwenCloud API, and the model is integrated into QwenWork, Qoder, and QwenStudio. A preview under the model ID qwen4-max-preview began November 11 and was scheduled to run through November 25; the general-release model ID qwen4-max launched on November 16, five days into the preview window [1][2][10].
Benchmarks: vendor-reported, one independent score
Alibaba published results on 16 benchmarks plus its internal Work360 test, claiming Qwen4-Max outperforms Qwen3.8-Max overall while using 35 percent fewer tokens at high reasoning effort and cutting average task cost by roughly 60 percent on Work360 [1][3]. The company says it beats Claude Opus 5.5, Kimi K3.1 Pro, and Muse Wonder Full on several benchmarks, including DeepSWE for coding (75.7 vs. 74.2 for Opus 5.5), but trails on Terminal-Bench 4.0 (48.5 vs. 59.6 for Opus 5.5 and 67.2 for Kimi K3.1) and ProgramBench (31.5 vs. 65.0 for Opus 5.5) [1][3]. Source [5] cites a different Anthropic model, calling it "Claude Fable 5.1" [5].
Artificial Analysis, an independent evaluator, gives Qwen4-Max an Intelligence Index score of 43, up from 34 for Qwen3.8-Max but behind Kimi K3.1 Pro at 46. Code Arena ranks it sixth in web development with a provisional score [1][3]. No independent replication of Alibaba's 16-benchmark table has been published; the Qwen 3.8 cycle showed independent verification typically takes weeks [1].
The open-weight promise: what "soon" means
Alibaba's launch post states smaller Qwen 4 models will be open-sourced, but provides no names, sizes, dates, or licenses [1][2][3][7][8]. At its Apsara Conference on September 22, the company named four tiers: Max (flagship), Plus (balanced multimodal), Flash (low-latency), and a 27B model for local deployment [1][3]. Only Max has shipped.
Precedent is mixed. The Qwen 3.8 generation followed a staged rollout: Qwen3.8-Max launched August 3, and Qwen3.8-27B open weights arrived around August 14 under Apache 2.0 — roughly ten days later [1][3]. The Qwen 3.7 generation, by contrast, shipped no open-weight models at all; Qwen3.7-Max and Qwen3.7-Plus remained API-only [1][3].
Licensing has varied by tier. 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 monthly revenue, and a separate license for inference-service or coding-assistant use above $50 million revenue over twelve months. Qwen3.8-Flash-Next, released August 28 as an "experimental preview of the architecture that will underpin Qwen4," uses the Qwen Community License 1.0, which requires a separate license for any inference service or coding/office assistant with no revenue threshold [1][3].
Architecture signals are real; the rest is timeline
The architecture is partly public. Qwen3.8-Flash-Next, downloadable since August 28, is a sparse MoE with 125 billion total parameters, 6 billion active per token, a 51-billion-parameter n-gram embedding component, native multimodality, and 262K context extensible to 1M. Qwen4-Max confirms the direction at flagship scale: 71B active of 2.4T backbone, 328B n-gram embedding, leaner activation, memory layer, multimodal by default [1][2][3].
Engine support is arriving before weights. vLLM has open pull requests implementing Qwen 4 sparse-attention, HyperConnection, and n-gram kernels, and merged upstream Qwen4-Exp model configs from Transformers — the same pattern that preceded Qwen 3.8 weights [1].
What remains unknown
- Which smaller models: Plus, Flash, 27B, or others — none confirmed by name in the November 16 announcement.
- Dates: "Soon" is the only timing given.
- Licenses: Could be Apache 2.0, custom, or community licenses per tier; Qwen 3.8 used all three.
- Specs for non-Max tiers: No parameter counts, active counts, context windows, or prices.
- Independent benchmarks: Only Artificial Analysis and Code Arena have published third-party scores so far.
- Qwen 4 Coder: Does not exist; the coding line remains Qwen3-Coder [1][3].
Safety claims
Alibaba says Qwen4-Max's training data was filtered to retain defensive cybersecurity knowledge (vulnerability detection, remediation) while removing offensive attack content. It also cites dedicated datasets and post-training enhancements for chemical, biological, radiological, and nuclear (CBRN) risk mitigation, and reports an improved secure-code-generation benchmark score of 14.7 percent (up from 8 percent for Qwen3.8-Max). Reuters notes these findings have not been independently verified [7].
Bottom line
Qwen4-Max is a real, shipped flagship with published specs, API pricing, and vendor benchmarks. The architecture advances — leaner MoE activation, large n-gram memory, native multimodality, 1M context — are demonstrable in the downloadable Flash-Next preview. But the open-weight commitment that defined Qwen's reputation through 3.8 is, for now, a promise without a model name, date, or license. The 3.8 precedent suggests weights could arrive in days; the 3.7 precedent shows Alibaba can ship a full generation without them. Developers planning to build on Qwen 4 open weights should watch the Qwen organization on Hugging Face and read the license on whichever tier appears first.
Sources
- Qwen 4: Release Date, Qwen4-Max Launch, and What's ... (yottalabs.ai)
- Qwen4-Max Preview: 1M Context, Cloud Only, No Weights (modelfit.io)
- Qwen 4 27B vs Qwen 3.8 27B: What to Expect from Qwen 4 (askthisguy.com)
- Alibaba says new flagship AI model cuts task costs by 60% ... (scmp.com)
- Alibaba Takes on Anthropic With New Qwen AI Model (pymnts.com)
- Qwen 4: Release Date, Models, and What We Know (1min.ai)
- Alibaba details new AI safeguards as Chinese developers step up test disclosures | Reuters (reuters.com)
- Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery - Alibaba Cloud Community (alibabacloud.com)
- Alibaba Unveils Roadmap on Full-Stack AI Strategy from Chips, Cloud Infrastructure, Models to Agents - Alibaba Cloud (alibabacloud.com)
- https://docs.qwencloud.com/changelog/models (docs.qwencloud.com)