Skip to content
NewsModel release

Alibaba promises open weights for Qwen3.8-Max this week: its first Max-class open-source release

· by Pondero Newsdesk

The short version

Alibaba committed to releasing Qwen3.8-Max weights on Hugging Face and ModelScope this week. The 2.4-trillion-parameter MoE API launched August 3 at $2/$6 per million tokens.

Alibaba promises open weights for Qwen3.8-Max this week: its first Max-class open-source release

Alibaba committed this week to publishing weights for Qwen3.8-Max on Hugging Face and ModelScope, making it the first Max-class Qwen model the company has ever open-sourced. The API went live on August 3 at $2 per 1 million input tokens and $6 per 1 million output tokens per the QwenCloud model page. The weights release covers both Qwen3.8-Max and a new companion Qwen3.8-27B dense model.

What

Qwen3.8-Max is a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active per token, with a 1-million-token context window (991K max input, 131K max output), per the Qwen team's announcement. Text, image, and video all feed into the same model call. A reasoning_effort flag lets callers trade inference cost for depth, choosing xhigh, medium, or low pass intensity.

Pricing on QwenCloud sits at $2.00 per million input tokens and $6.00 per million output tokens at cache miss, dropping to $0.25 per million on implicit cache reads, per Developers Digest's verified price table. Vercel added it to AI Gateway on August 2 at provider pricing with no markup, giving teams a zero-overhead access path alongside the native QwenCloud API.

Benchmarks are vendor-stated using Alibaba's own harnesses on most runs. The Qwen announcement puts Qwen3.8-Max at 93.0 on PaperBench, 86.1 on OSWorld-Verified, and 88.0 on RealWorldQA. On SWE-bench Pro, the model scored 67.7, trailing Fable 5 at 80.0. Cross-lab comparisons used different harnesses per vendor, so treat the numbers as directional rather than settled.

The three prior Max-class Qwen models, Qwen3-Max, Qwen3.6-Max, and Qwen3.7-Max, launched hosted-only and stayed that way. None of them carried an open-weight commitment; Alibaba simply kept the flagship tier closed while smaller lines shipped permissively, per Digital Applied's closed-flagship analysis. Qwen3.8 breaks that pattern with a dated commitment for the first time.

The 2.4-trillion-parameter size creates a real serving constraint. MoE routing activates only 95 billion parameters per token, but the full expert set must remain resident in accelerator memory. Digital Applied's hardware breakdown estimates the weight set at roughly 1.2TB at four-bit quantization before key-value cache or batch headroom. The 27B companion sits in a very different class: a single high-end GPU handles it at quantized precision, making it the artifact most teams can deploy without new hardware.

Why it matters

Pricing is the immediate lever for operators using the API today. At $6 per million output tokens, Qwen3.8-Max costs less than half of Kimi K3's $15 output rate per Developers Digest's price table, while operating at broadly comparable scale (2.4T vs 2.8T parameters, both MoE). For agent loops that re-read large codebases repeatedly, cache reads at $0.25 per million input tokens cut effective cost sharply across long sessions.

The open-weights window matters most for regulated industries, on-prem deployments, and teams planning to fine-tune. If the weights land this week, operators gain a 2.4-trillion-parameter frontier option they can run on their own infrastructure, move data without it leaving their boundary, and adapt without Alibaba's consent. The nearest comparable drop was Kimi K3: Moonshot attached an open-weight commitment on July 17 and published the weights on July 27, per Digital Applied's base-rate analysis. The Qwen3.8 window opened today.

One decision still pending: the license. Qwen's smaller lines (Qwen3.6-27B) shipped under Apache 2.0 on Hugging Face. That precedent is real, but the license for Qwen3.8-Max has not been disclosed as of August 10. Teams planning commercial redistribution or fine-tune-and-serve deployments should read the LICENSE file in the official Hugging Face repository before writing deployment tickets, not afterward.

What to watch next

The first event to track is whether a repository appears under the Qwen organization at huggingface.co/Qwen this week. Separately, the Qwen3.8-27B carries no published specs beyond the parameter count. No benchmark table, context figure, or license has been confirmed for it. If the Max license turns out to be Apache 2.0, it would extend Alibaba's open-weight pattern to the frontier tier for the first time, and shift the negotiating position of every operator currently locked into a hosted API for that scale.

Sources