Alibaba Open-Sources Qwen3.8-2.4T-A95B: What Developers Need to Know

On August 12, 2026, Alibaba open-sourced the Qwen3.8-2.4T-A95B model — the first time it has released open weights for a "Max"-class flagship. It is the core of the Qwen3.8-Max system that debuted on August 3. Below is the full breakdown.

Key Specifications

Metric	Value
Total parameters	2.4 trillion
Active parameters per token	~95 billion
Architecture	Sparse Mixture-of-Experts (MoE), 92 layers, 512 experts (10 routed + 1 shared active per token)
Attention	Hybrid Gated DeltaNet + Gated Attention
Native context	262,000 tokens, extendable to ~1,010,000 tokens
Max input	991K tokens (983K with thinking enabled)
Max output	131K tokens
Vocab size	248,320
Precision	BF16 and FP8 weights available
Modality (open weights)	Text-only input/output
License	Custom revenue-share license; MaaS/AI Work Assistants with >$50M annual revenue must negotiate separately

The open-weights checkpoint was published on Hugging Face and ModelScope .

Benchmark Performance (Alibaba's Internal Figures)

Alibaba reported Qwen3.8-Max as "second only to Fable 5" among tested systems . Key published benchmarks :

Benchmark	Score
Terminal-Bench 2.1	86.6
GPQA Diamond	92.6
SWE-bench Pro	67.7 (vs. Fable 5's 80.0)
FrontierSWE	73.8 (vs. Fable 5's ~80+)

Important caveat: These are Alibaba's internal runs. No independent benchmarks from Artificial Analysis or LMArena were available as of August 12 . On core coding tasks, Qwen3.8-Max trails Anthropic's Fable 5 by a meaningful margin .

Commercial Pricing (QwenCloud API, Full Qwen3.8-Max)

Service	Price
Input tokens	$2.00 per million tokens
Output tokens	$6.00 per million tokens
Implicit cache reads	$0.25 per million tokens
Explicit cache creation	$2.50 / $0.17 read

The open-weights model itself is free to download, but the license restricts commercial MaaS usage above a $50M revenue threshold, pushing high-revenue users toward a paid commercial agreement .

Limitations Compared to the Full Paid Version (Qwen3.8-Max API)

The open-weights Qwen3.8-2.4T-A95B is a stripped-down subset of the full Qwen3.8-Max cloud model:

• No vision input — text-only, whereas the API supports images, video, and documents .
• No native 1M context — the open weights ship with 262K native context (expandable), while the API natively supports 1M tokens .
• No default non-thinking mode — the API can run in non-thinking mode without manual configuration; the open model requires users to handle this .
• No built-in official tools — the API includes integrated tooling (search, code execution, etc.) that the open weights lack .
• Missing companion models — the promised Qwen3.8-27B smaller model did not ship alongside the open weights .
• License restrictions — any MaaS or AI Work Assistant provider exceeding $50M in revenue in a 12-month period must negotiate a separate commercial license. The license defines MaaS revenue broadly, including API calls, fine-tuning, agent subscriptions, and inference credits from monthly plans .

Broader Strategic Context in Chinese AI Competition

• Direct response to rivals: The July 2026 open-weight launch of Moonshot's Kimi K3 (a multimodal 1M-context model) put pressure on Alibaba to open its own flagship . Qwen3.8-2.4T-A95B arrived roughly six weeks later.
• "Open but gated" platform play: The revenue-contingent license ($50M threshold) is widely interpreted as a moat-building strategy — Alibaba open-sources to drive ecosystem adoption and developer mindshare, but caps free commercial use to protect its cloud API revenue (QwenCloud) and steer enterprise customers toward paid contracts .
• Scaling race within China: The model is 7× larger than Qwen3.5 (released February 2026) . Alibaba framed Qwen3.8-Max as "second only to Fable 5," positioning itself as China's strongest contender against U.S. frontier labs, just weeks after the U.S. briefly restricted Fable 5's export .
• Hardware/self-hosting barrier: A 2.4T-parameter model at 4-bit precision requires roughly 1.2 TB of GPU memory for weights alone , meaning self-hosting is viable only for well-resourced organizations — further concentrating deployment on Alibaba Cloud.
• Portfolio strategy: The open-weights release runs alongside Alibaba's larger "Qwen3.8 family" strategy, with smaller models (27B expected), specialized variants for coding and agent use, and vision-turbo versions that keep multimodal capabilities on the paid API tier .