Qwen 3.8 Max: Alibaba releases an open-weights model with 2.4 trillion parameters

On August 3, 2026, Alibaba announced Qwen 3.8 Max — the new flagship of its Qwen lineup. On paper, the numbers are staggering: 2.4 trillion parameters in total, with 95 billion active per token (Mixture-of-Experts architecture), 1 million context tokens, native multimodal (text, image, video input). The weights will be released open source about a week after the announcement.

It is the second largest open-weights model in the world, behind Kimi K3 from Moonshot (2.8 T), released the same week. China continues to flood the open model market.

The architecture: accelerated sparse MoE

Like Kimi K3 and DeepSeek V4 Pro, Qwen 3.8 Max relies on the Mixture of Experts pattern:

• 2.4 trillion total parameters stored (weights)
• Only 95 billion are activated to process each token
• Result: inference cost comparable to a dense model of ~100 B, specialization capacity of a giant

This architecture allows a colossal brain to fit on a reasonable infrastructure — a sine qua non for open source democratization.

Native multimodality

Qwen 3.8 Max accepts input:

• Text
• Images (understanding and analysis)
• Video (temporal understanding, sequence summarization)

Output is text-only for now — no image or video generation (Alibaba has dedicated models for that: Wan for video, notably).

The open-weights landscape of 2026

The consolidated table of the current top 5 open-weights:

1. Kimi K3 (Moonshot, China) — 2.8 T
2. Qwen 3.8 Max (Alibaba, China) — 2.4 T
3. DeepSeek V4 Pro (DeepSeek, China) — 1.6 T
4. Llama 5 (Meta, USA) — 1.2 T
5. Mistral Large 3 (Mistral, France) — 700 B

Three out of five are Chinese, including the top two. Meta and Mistral hold their ground on the Western side. The gap with closed models (Claude, GPT, Gemini) narrows benchmark after benchmark.

A year after DeepSeek V3, China is no longer catching up: it sets the pace for open models.

What it changes concretely

Three practical shifts:

1. In-house fine-tuning becomes accessible again. French and European companies hesitating between fine-tuning a Llama 4 vs paying Anthropic now have a third choice: Qwen 3.8 Max with a permissive license and top-tier performance.
2. European GPU hosts (Scaleway, OVHcloud, IONOS AI) will offer Qwen 3.8 Max in the coming weeks. Good news for sovereignty: hosting a Chinese model on European soil without US API dependency.
3. Anthropic and OpenAI see their pricing challenged. A model that rivals in quality, runs at a tenth of the cost, with no commitment — it pushes all prices down.

The nuance

Beware of three points:

• Official benchmarks published by Alibaba are favorable — as always. Independent evaluations (LMSYS, LiveBench, SWE-Bench) are expected in the coming weeks to decide.
• “Open-weights” ≠ “open-source”. Training data, pipeline code, internal evaluations are not published. We get the final model, not the recipe.
• Geopolitical considerations. A Chinese model hosted by a European operator remains a Chinese model — possible biases on sensitive topics (Taiwan, Tibet, historical events). Regulated companies (finance, defense, health) will need to assess.

What to watch

Two milestones:

• The effective release of the weights (announced for the week of August 10)
• The independent third-party evaluations — LMSYS, LiveBench, SWE-Bench Verified — expected by mid-September

If Qwen 3.8 Max lives up to its promises, it's one of the three or four models to integrate by default into any enterprise AI stack in 2026-2027. And a further reminder that the center of gravity of open source is inexorably shifting from west to east.