NRamosDev · Blog
Qwen 3.8 Max: the 2.4-trillion giant that promises open weights
#qwen#ai#llm#open-weights
Why this launch matters
On August 3, 2026, Alibaba released Qwen 3.8 Max, a 2.4-trillion parameter model designed to lead the agentic intelligence era. For the first time, the Chinese firm has promised to open the weights of its flagship Max-class model.
Architecture: the beast within
- Mixture of Experts: only 95 billion parameters are active per token out of 2.4 trillion total.
- Colossal context: 983K input tokens (nearly a million) and a 128K output limit.
- Natively multimodal: video, images, documents and text in one pipeline.
- Pricing: $2 per million input tokens, $6 for output — a clear reduction vs Qwen 3.7 Max.
Benchmarks
- OSWorld Verified: 86.1 — leading, ahead of Claude Fable 5 and GPT-5.6 Sol.
- PaperBench: 93.0 — dominant.
- Terminal-Bench 2.1: 86.6 — above Fable 5, below GPT-5.6’s 88.8.
- IFBench: 82.8 — leading.
- SWE-Bench Pro: 67.7 — its weak spot, losing to Claude Fable 5’s 80.0.
The Chinese arms race
- Kimi K3 (Moonshot AI): 2.8 trillion total, 104 billion active. Weights released in MXFP4 (1.56 TB on disk).
- DeepSeek V4 and GLM-5.2 (Zhipu AI): also in the fight.
Open weights: the promise that keeps waiting
On August 2, Alibaba said weights would be released the following week. Today, no trace of the model on Hugging Face.
Running 2.4 trillion parameters locally is fantasy: 4-bit quantization needs 1.2 TB of VRAM. The real local path will be the upcoming Qwen 3.8 27B.
This article accompanies the video on the nramosdev channel. Subscribe for more AI analysis in Spanish.