Introducing Qwen3.8-Max open weights

For most of the past two years, "open weights" and "frontier class" lived on different shelves. Qwen's flagship tier - the "Max" line - was always API-only, no matter what the smaller siblings did. That changed last week.

Today, we are putting the largest open-weight release in history into context for healthcare teams. Qwen3.8-Max's open weights, published August 12, 2026, are the first time a Qwen-Max-class model has shipped as a downloadable checkpoint - a 2.4 trillion parameter Mixture-of-Experts model, with a custom license and some real caveats - and we take a close look at what is actually inside, what it costs to serve, and where it fits into a HIPAA-compliant stack under our design-partner program.

What is Qwen3.8-Max (Qwen3.8-2.4T-A95B)?

The release is staged in two steps. Alibaba announced Qwen3.8-Max on August 3, 2026 with the API and a statement that open weight release would follow; the actual weights went live on Hugging Face around August 12, 2026 under the model ID Qwen/Qwen3.8-2.4T-A95B. Both dates matter, and we list both because the gap between announcement and downloadable checkpoint has been the whole story of 2026 open-weight releases.

The headline specs:

• 2.4 trillion parameters total, 95 billion active per token, a Mixture-of-Experts model with 512 experts, 10 routed + 1 shared per layer.
• A 92-layer stack combining Gated DeltaNet (linear attention) and Gated Attention in a repeating three-plus-one pattern, which is the architectural foundation carried over from the Qwen3.5 series.
• 262,144 token context natively, extensible to about 1,010,000 tokens on the open checkpoint.
• Text-only on the open weights, thinking-only at inference. Multimodal input and a non-thinking mode are features of the hosted Qwen3.8-Max API, not of this checkpoint. Every response carries a reasoning block by default, and reasoning_effort supports xhigh (default), medium, and low.

Two things to be clear about. First, this is a post-trained checkpoint aimed at teams running vLLM or SGLang on serious hardware - the model card explicitly frames it that way, not as a hobbyist release. Second, the license is not the Apache 2.0 Qwen is known for. The card lists a custom qwen3.8-max license, and reading the actual terms matters here: free commercial use carries two conditions, not an outright ban. One, if a commercial product using these weights crosses 100M monthly active users or $20M monthly revenue, the product UI must prominently display the Qwen model name. Two, and more relevant for an inference service: anyone operating a "Model as a Service" (hosted API access for third parties) whose aggregate affiliate-group revenue exceeds $50M across any rolling 12 months must obtain a separate license from Qwen before serving the model commercially. Below that threshold, hosting is permitted on the stock license - which is why we can bring this checkpoint into our HIPAA-compliant catalog without a bespoke arrangement today.

Performance profile: what it's good at vs. what it's not

What Qwen3.8-Max excels at

• Agentic terminal work at frontier level. On the vendor's reported table (Terminus 2 harness, max effort, temperature 1.0, top_p 0.95), it posts 86.6 on Terminal-Bench 2.1, within a point of Claude Opus-4.8 and Fable-5 (both 84.6), behind GPT-5.6 Sol at 88.8, and well ahead of Qwen3.7-Max at 74.5.
• Research-style reasoning. On PaperBench it posts the highest figure in Qwen's table at 93.0, ahead of every baseline reported - long, multi-step scientific reasoning over documents, which is the shape of most clinical-evidence synthesis work.
• Professional and legal reasoning. On Qwen's own PLawBench at 73.2 and HealthBench at 60.2, it leads every model in its comparison table, including Opus-4.8 - numbers worth flagging because they are vendor-judged and in-domain for exactly the workloads our readers run.
• Structured reasoning depth control. reasoning_effort (xhigh / medium / low) plus preserve_thinking gives operators a real dial, not a binary thinking toggle.

Limitations to keep in mind

• Text-only on the open checkpoint. Vision input and non-thinking mode require the hosted Qwen3.8-Max API release, not the weights. If a workflow needs to see an image today, run a multimodal model alongside this one.
• Hardware, not hype. The full BF16 checkpoint is several terabytes on disk, and public serving recipes from the community target GB300 NVL72-class racks. If you cannot operate that hardware, self-hosting is not realistic, and the practical deployment path is a managed API - which is the constraint our design-partner program exists around.
• Custom license. "Open weights" does not mean "Apache 2.0". Before deploying this as part of any commercial service, the exact qwen3.8-max license terms need legal review on your side, not ours.
• Always-on reasoning cost. Every turn runs a reasoning block by default on the open checkpoint. For high-volume transactional workloads (eligibility checks, form extraction, status lookups), that verbosity is real money. For high-stakes reasoning, it is exactly what you want; for high-volume summarization, it is not.
• The rest of the Qwen3.8 family is stronger for most teams. Their Qwen3.8-27B release, a few days later, gives away the dense, consumer-runnable half of this family under Apache 2.0, and we think it is the more practical entry point for most deployments.

The benchmarks: how Qwen3.8-Max compares

Benchmarks are from the official Qwen3.8-2.4T-A95B model card (vendor-reported, Terminus 2 harness where noted; baselines are described by Qwen as "best published score across harnesses" and use Opus-4.8 and Fable 5 from Terminus 2 on Artificial Analysis, GPT-5.6 Sol from OpenAI's Codex preview page).

The pattern is consistent with other frontier open releases: it trades blows with closed frontier models rather than beating them outright, and the gaps it loses - SWE-bench Pro and DeepSWE specifically - are the same gaps every open model in August 2026 is losing.

Why run Qwen3.8-Max via our HIPAA-compliant API?

• Enterprise-grade privacy. Available under our HIPAA-compliant design-partner program with a signed BAA, end-to-end data encryption, and strict zero-data-retention on prompts and outputs. Your patient data is never used for training. General availability is coming soon.
• Uncompromised performance. We handle the multi-node serving hardware, the bf16 weight distribution, and the reasoning_effort / preserve_thinking plumbing - you get 2.4T flagship reasoning without buying a rack.
• Transparent pricing. Qwen3.8-Max runs at $2.00 per 1M input tokens, $0.25 per 1M cached input, and $6.00 per 1M output tokens on our serving layer, scaling with the active 95B parameters rather than the total stack - a structurally cheaper posture than hosted frontier APIs on Azure OpenAI or AWS Bedrock.

Ready to put 2.4T-parameter open reasoning behind your BAA? Join the waitlist now.