Qwen 3.5 Launches on New Year’s Eve as DeepSeek Only Sends a Holiday Greeting

Alibaba’s Qwen team quietly released the open‑source Qwen 3.5 series on New Year’s Eve, announcing the first open‑weight model Qwen3.5‑397B‑A17B on Hugging Face.

397B Body, 17B Soul

397B : total parameters close to 400 billion, a true giant.

A17B : only 17 billion activation parameters are used during inference.

The model combines Hybrid Linear Attention with Sparse MoE , giving it the knowledge of a 400 billion‑parameter brain while activating just 17 billion neurons for a specific query.

Result: decoding throughput is 8.6–19× higher than the previous Qwen3‑Max, while maintaining strong reasoning ability.

Performance vs. Closed Models

Agent ability (BFCL V4) : 72.9, surpassing GPT‑5.2 (63.1) and Claude Opus 4.5 (67.7).

Code ability (SWE‑bench Verified) : 76.4, slightly below GPT‑5.2 (80.0) but ahead of Qwen3‑Max‑Thinking.

Native multimodal (Video‑MME) : 87.5, showing excellent video‑understanding thanks to its built‑in visual module.

Scaling Law Still Holds

The team published a scaling‑law chart demonstrating that the hybrid architecture continues to improve performance steadily as model size grows.

Apache 2.0 License

The model is released under Apache 2.0, meaning it is completely free for commercial use.

Front‑end Demo: One Prompt Generates a 3D Car Game

In a showcase, a single natural‑language command produced a fully runnable 3D racing game.

Real‑time scoring : scores displayed at the top‑left.

Lap counter : tracks completed laps.

Timer : millisecond‑precision timing.

Speedometer : dynamic speed readout at the bottom‑right.

Integration with OpenClaw: Search, Think, Report

"Help me search for AI models released in the past month, compile a report, and generate a PDF."

Automatic search : calls a search tool to retrieve latest model releases (Claude Sonnet 5, Kimi K2.5, GPT‑5.3‑Codex‑Spark, DeepSeek V3.2, Gemini 3, etc.).

Information integration : structures the results into a report with model name, release date, key features, and benchmark data.

PDF generation : uses a document‑generation tool to export a 70 KB PDF.

The final PDF includes detailed model information, professional layout, and data visualizations.

Multimodal Agent: See → Think → Search → Report

Qwen 3.5‑Plus combines native multimodal vision with agent capabilities.

In a demo, a user uploads a screenshot from the movie “Gone with the Wind” and asks, “This film looks familiar—can you introduce it?”

Step 1: Visual analysis of the image.

Step 2: Search to verify information.

Step 3: Generate a concise report.

The closed loop “see → think → search → generate report” showcases strong multimodal reasoning and tool‑calling.

Ecosystem Blossoms: Rapid Multi‑Platform Support

OpenRouter – API Aggregation Platform

OpenRouter integrated Qwen 3.5 within six hours, allowing developers to call the model via a unified API without self‑hosting.

Poe – Conversational Platform

Quora’s Poe added Qwen 3.5‑Plus (1 M‑token context) alongside GPT‑5.2‑Codex, Gemini‑3‑Pro, Claude‑Opus‑4.6, and Kimi‑K2.5, highlighting its competitive capability.

Ollama – Local Deployment Tool

Run the model locally with a single command: ollama run qwen3.5:cloud The model is also accessible via Ollama’s cloud service for immediate testing.

NVIDIA – Hardware Giant Support

NVIDIA Build : free build via build.nvidia.com/qwen/qwen3.5-3....

NVIDIA NeMo : supports download and custom deployment from the GitHub repository.

NVIDIA’s backing adds hardware‑level optimization and eases enterprise adoption.

Overall, Qwen 3.5’s release is more than a model update; it sparks a collective celebration across the open‑source AI ecosystem, offering extreme inference speed, leading agent ability, and native multimodal interaction.