Qwen3.5: The Open Frontier Model with 397B Parameters
Alibaba just released Qwen3.5, and it’s a statement: 397 billion parameters, open weights, native multimodal capabilities, and support for 201 languages.
The frontier is open.
The Numbers
- 397B total parameters, only 17B active (Mixture-of-Experts)
- 201 languages and dialects
- Native multimodal (vision + language from the start)
- Open weights on HuggingFace
The MoE architecture means you get near-400B capability with inference costs closer to a 20B model. That’s the efficiency breakthrough that makes this practical.
What’s New in 3.5
Unified Vision-Language Foundation — Not a language model with vision bolted on. Early fusion training on multimodal tokens from the start. Achieves parity with Qwen3 on text while matching Qwen3-VL on vision tasks.
Gated Delta Networks + MoE — A hybrid architecture combining linear attention (Gated DeltaNets) with sparse MoE. High throughput, low latency, efficient inference.
Massive RL Scale — Trained with reinforcement learning across “million-agent environments” with progressively complex tasks. This is how you get robust real-world performance.
Global Language Coverage — 201 languages isn’t a checkbox feature. It’s inclusive deployment with actual cultural and regional understanding.
Architecture Details
Total Parameters: 397B
Active Parameters: 17B
Hidden Dimension: 4096
Layers: 60
Token Vocab: 248,320
Layer Pattern (15x):
3x (Gated DeltaNet → MoE)
1x (Gated Attention → MoE)
The Gated DeltaNet layers handle efficient sequence processing, while periodic Gated Attention layers provide full attention when needed. Smart hybrid design.
Capabilities
- Coding — Competitive with frontier closed models
- Reasoning — Strong on math, logic, planning
- GUI Agents — Can interact with interfaces
- Video Understanding — Native temporal reasoning
- Multilingual — Actually works across languages
Getting Started
Available through:
- HuggingFace:
Qwen/Qwen3.5-397B-A17B - vLLM, SGLang: Direct compatibility
- Alibaba Cloud: Managed API with 1M context, built-in tools
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3.5-397B-A17B",
device_map="auto"
)
Why This Matters
Every few months, someone claims “open source is catching up.” Qwen3.5 might actually deliver on that promise.
The combination of:
- Frontier-scale parameters
- Efficient MoE inference
- True multimodal capabilities
- Global language support
- Open weights
…makes this immediately useful for production applications that previously required closed APIs.
For teams building multilingual products, agents that need vision, or applications requiring on-premise deployment—this is the new baseline to evaluate against.
My Take
We’re past the point where “open vs closed” is a capability gap. It’s now about trade-offs: latency, cost, control, compliance. Qwen3.5 gives you a frontier-capable model you can actually run and modify.
The 201-language support is underrated. Most models are English-first with other languages as afterthoughts. Qwen has been genuinely multilingual from the start, and 3.5 extends that lead.
Links: