Pull to refresh
Logo
Alibaba releases Qwen3.8-Max as open weights, undercutting US model pricing

Alibaba releases Qwen3.8-Max as open weights, undercutting US model pricing

New Capabilities

The weights are out, with new licensing strings attached, and DeepSeek already undercuts the price

August 12th, 2026: Qwen3.8-Max weights ship, but with new restrictions

Overview

Updated Aug 14

Alibaba's Qwen3.8-Max weights landed on Hugging Face on August 12, nine days after the API launch. The download is smaller than promised: it skips the 1-million-token context and image input of the hosted model, and the 27-billion-parameter version has not shipped.

The release also carries a new license, not the fully open Apache 2.0 terms Alibaba used for earlier Qwen models. Businesses earning more than $20 million a month must credit Alibaba, and AI-assistant companies above $50 million in trailing revenue need a separate paid license. Days later, DeepSeek shipped its own V4 Pro model at $0.44 and $0.87 per million tokens, undercutting Qwen's price again.

Why it matters

A frontier-grade AI you can download for free, priced below US rivals, puts advanced capability in more hands and squeezes closed-model economics.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

2.4T
Total parameters
Sparse mixture-of-experts design; the largest model Alibaba has shipped.
95B
Active parameters per query
Only a fraction of the 2.4 trillion fire on any single request, cutting compute cost.
1M tokens
Context window (hosted API)
The paid API reads roughly 750,000 words in one prompt. The open-weight download published August 12 leaves this out, along with image input.
$2 / $6
Qwen3.8-Max price per million tokens (in / out)
OpenAI's flagship GPT-5.6 Sol charges $5 in and $30 out. Qwen undercuts it by about 60% and 80%.
$0.44 / $0.87
DeepSeek V4 Pro price per million tokens (in / out)
DeepSeek's flagship, shipped August 12, undercuts Qwen3.8-Max's own price by roughly 78% on input and 86% on output.
~7%
Alibaba Hong Kong share jump
Hong Kong stock rose about 7% and the New York listing about 4.5% on launch day, August 3.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Sign up to generate historical perspectives on this story.

People Involved

Organizations Involved

Timeline

January 2025 August 2026

8 events Latest: August 12th, 2026 · 1 month ago
Tap a bar to jump to that date
  1. Qwen3.8-Max weights ship, but with new restrictions

    Latest Release

    Alibaba published the Qwen3.8-2.4T-A95B checkpoint on Hugging Face under a new license, not the Apache 2.0 terms of earlier Qwen models. The download skips the hosted model's 1-million-token context and image input.

  2. DeepSeek ships V4 Pro, undercuts Qwen's price again

    Release

    DeepSeek quietly moved V4 Pro out of preview at $0.44 per million input tokens and $0.87 per million output tokens, well below Qwen3.8-Max's $2 and $6.

  3. Qwen3.8-Max goes live

    Release

    Alibaba makes its most capable model widely accessible, undercuts OpenAI's flagship GPT-5.6 on price, and promises open weights within a week. Shares jump.

  4. Alibaba previews Qwen3.8-Max

    Announcement

    Days after Kimi K3, Alibaba teases a 2.4-trillion-parameter multimodal model bound for open release.

  5. Moonshot launches Kimi K3

    Release

    A 2.8-trillion-parameter open-weight model sets a new bar for scale and agent-focused coding.

  6. Qwen leaders depart

    Personnel

    Technical lead Lin Junyang and other senior researchers leave Alibaba; Eddie Wu takes direct control of AI strategy.

  7. Alibaba ships Qwen3-Max, closed

    Release

    Alibaba's first trillion-parameter model launches as API-only, with no downloadable weights.

  8. DeepSeek R1 jolts the market

    Milestone

    A low-cost Chinese reasoning model matches Western frontier results and triggers a sharp US tech-stock selloff.

Scenarios

1

Qwen3.8-Max weights land on Hugging Face on schedule

Likely Resolves by Sep 3, 2026

Discussed by: SCMP, MarkTechPost, MLQ News

Alibaba said downloadable weights would follow within about a week on Hugging Face and ModelScope. The team has a track record of shipping promised open weights, as Moonshot did with Kimi K3. If a public model card and downloadable checkpoint appear on either platform, the pledge holds.

2

Qwen3.8-Max reaches the LMArena top three

Possible Resolves by End of 2026

Discussed by: Forkast, Yahoo Finance, LMArena watchers

Alibaba claims the model rivals Anthropic's Claude Fable 5 and leads on coding and multimodal tests, while trailing on general reasoning. Public leaderboards will test that claim against live user votes. A top-three text-arena finish would confirm frontier standing; a lower placing would suggest the benchmarks flattered it.

3

A major US lab ships an open-weight frontier model in response

Uncertain Resolves by Q2 2027

Discussed by: The New Stack, Nathan Lambert (Interconnects)

Chinese labs now dominate the open-weight tier. Cheap, downloadable frontier models pressure OpenAI and Anthropic's closed, paid strategy. One response would be a US lab releasing its own frontier-class open-weight model to defend developer mindshare. So far Western majors have mostly held their best models back.

4

US government moves to restrict Chinese open-weight models

Unlikely Resolves by Jan 20, 2027

Discussed by: Policy analysts cited by TechRepublic and Forbes

Wide US adoption of downloadable Chinese models raises security and dependence concerns in Washington. A formal response could bar federal use or hosting of models like Qwen3.8-Max. No such rule exists yet, and open weights are hard to control once distributed.

5

Alibaba loosens Qwen3.8-Max's license after developer pushback

Uncertain Resolves by Nov 30, 2026

Discussed by: Forkast News, Hugging Face community discussion

Alibaba's new license breaks from the fully open Apache 2.0 terms of earlier Qwen releases, adding revenue-based fees for large deployers. If enterprise users or open-source advocates push back, Alibaba could revise the terms, as it has adjusted licensing before.

6

A top-ranked model prices output below $1 per million tokens

Likely Resolves by Oct 31, 2026

Discussed by: OpenRouter pricing data, Artificial Analysis

DeepSeek's V4 Pro already prices output at $0.87 per million tokens. If Alibaba, Moonshot, or another major lab matches or beats that on a top-ranked model, frontier output pricing would fall below $1 per million tokens for the first time.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

November 2007 to 2008

Google open-sources Android (2007-2008)

Google released the Android mobile operating system for free instead of charging device makers. It gave away the platform to expand the market for its search and services.

Then

Handset makers adopted Android quickly, spreading it across low- and high-end phones.

Now

Android became the world's most-used mobile OS and fed Google's core ad business for years.

Why this matters now

Alibaba gives away the model to sell the complement: cloud compute. Free weights can be a route to platform control, not charity.

July 2023

Meta releases Llama 2 open weights (July 2023)

Meta published Llama 2 with weights free for most commercial use. A leading US company chose to give away a strong model rather than sell access to it.

Then

Developers built a large ecosystem of fine-tuned variants on top of the free base model.

Now

Open weights became a real strategy, pressuring closed labs and seeding the tools Chinese labs later scaled.

Why this matters now

Alibaba is running Meta's playbook at frontier scale, using free models to draw developers toward its paid cloud.

January 2025

DeepSeek R1 market shock (January 2025)

A little-known Hangzhou lab released R1, a reasoning model that matched top US systems at a fraction of the training cost. Investors realized frontier AI might not require the spending they had assumed.

Then

US tech stocks sold off sharply, with chipmaker Nvidia among the hardest hit in a single session.

Now

The release reset the AI contest around open weights and price, the terms Alibaba and Moonshot compete on today.

Why this matters now

Qwen3.8-Max extends the pattern R1 started: Chinese labs matching Western capability, then competing on openness and cost.

Sources

(14)