On March 2, 2026, Alibaba open-sourced the Qwen3.5 small-scale model series. The 9B version achieved 81.7 on GPQA Diamond, surpassing OpenAI's GPT-OSS-120B (71.5). With a 13.5x parameter gap, the small model won.
Apache 2.0 license means both code and weights are available for commercial use. It runs with a single Ollama command and can be deployed on standard laptops.
Figure 1: Qwen3.5 Small Model Performance Comparison (Source: GitHub README)
1. Qwen3.5 Small-Scale Model Series
On March 2, 2026, Alibaba Qwen team open-sourced four Qwen3.5 small-scale models: Qwen3.5-0.8B, 2B, 4B, and 9B.
This is not a "shrunk version." This series uses native multimodal training with the latest model architecture.
Figure 2: Qwen3.5 Middle-Sized Model Performance (Source: GitHub README)
Model Positioning:
| Model | Positioning | Features | Use Cases |
|---|---|---|---|
| 0.8B/2B | Edge Device Choice | Extremely small, ultra-fast inference | Mobile devices, IoT, real-time interaction |
| 4B | Lightweight Agent | Multimodal base | Agent core |
| 9B | Compact Size, Cross-Level Performance | Competes with 120B | Server-side, memory-constrained |
0.8B and 2B are suitable for mobile devices and IoT edge deployments. 4B is ideal for lightweight agents. 9B is perfect for server-side deployment with excellent cost-effectiveness.
2. 9B vs 120B: Benchmark Data
GPQA Diamond benchmark results:
| Model | GPQA Diamond | Parameters | Approach |
|---|---|---|---|
| Qwen3.5-9B | 81.7 | 9B | End-to-End |
| GPT-OSS-120B | 71.5 | 120B | End-to-End |
Qwen3.5-9B scores 10.2 points higher than 120B.
VentureBeat's headline was direct: "Alibaba's small, open source Qwen3.5-9B beats OpenAI's gpt-oss-120B and can run on standard laptops."
What does "can run on standard laptops" mean? The 9B model uses about 4-5GB VRAM. RTX 3090, A10, or even high-end laptop GPUs can run it. No need for datacenter-grade GPUs like A100 or H100.
Previously, running a 120B model required at least 8 A100s. Now the 9B model works on a single card. The cost difference is orders of magnitude.
3. Technical Highlights: Why Can Small Models Win?
Qwen3.5 is not "distillation" or "pruning." There are several technical breakthroughs:
1. Unified Vision-Language Foundation
Early fusion training with trillions of multimodal tokens. Qwen3.5 surpasses Qwen3-VL models in reasoning, encoding, agent capabilities, and multimodal understanding.
Figure 3: Qwen3.5 Flagship Model Performance Comparison (Source: GitHub README)
2. Efficient Hybrid Architecture
Gated Delta Networks combined with sparse MoE (Mixture-of-Experts). High-throughput inference with low latency.
Qwen3.5-397B-A17B has 397B total parameters but only activates 17B per forward pass. Qwen3.5-9B doesn't disclose its MoE configuration but inherits the same architectural philosophy.
3. Scalable RL Generalization
Scaling reinforcement learning in millions of agent environments. Not optimization for specific benchmarks, but real-world adaptability.
4. Global Language Coverage
Expanded from 119 to 201 languages. Vocabulary grew from 150k to 250k, improving encoding/decoding efficiency by 10-60%.
4. Practical Deployment: One Command
How simple is deploying Qwen3.5-9B? One Ollama command:
ollama run qwen3.5:9b
That's it.
Using transformers:
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
model = Qwen3VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen3.5-9B",
torch_dtype=torch.bfloat16,
device_map="auto",
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-9B")
VRAM Usage:
- bfloat16 precision: ~4-5GB
- int8 quantization: ~2-3GB
- int4 quantization: ~1-2GB
Inference Speed (Single RTX 3090):
- Generation speed: ~30-50 tokens/second
- First token latency: <100ms
Comparison with 120B model:
- VRAM usage: ~240GB (bfloat16)
- Requires: 8 A100s (80GB each)
- Inference speed: ~5-10 tokens/second
The difference is clear.
5. Selection Guide: How to Choose 0.8B/2B/4B/9B?
| Requirement | Recommended Model | Reason |
|---|---|---|
| Mobile Deployment | 0.8B/2B | Extremely small, ultra-fast inference |
| IoT Edge Devices | 0.8B/2B | Low resource consumption |
| Lightweight Agent | 4B | Balanced performance and resources |
| Server General Purpose | 9B | Best cost-effectiveness |
| Memory <4GB | 0.8B/2B | Minimum resource requirements |
| Memory 4-8GB | 4B/9B | Medium resource requirements |
| Maximum Performance | 9B | 接近 120B performance |
Recommendations:
- Ample memory (≥8GB): Go straight for 9B
- Mobile development: Choose 2B
- Agent development: 4B is the sweet spot
6. Conclusion: The Era of Small-Scale Models
Qwen3.5-9B open-source release marks a new trend: small-scale models are no longer a "compromise" but a "choice."
Previously, performance = parameters. The fact that 9B exceeds 120B tells us: architecture optimization > piling parameters.
This is good news for developers. Previously limited to cloud API calls, now can deploy locally. Previously worried about data privacy, now can run completely offline. Previously too expensive, now works on a single card.
Resources and References
- GitHub: github.com/QwenLM/Qwen3.5
- ModelScope: modelscope.cn/collections/Qwen/Qwen35
- HuggingFace: huggingface.co/collections/Qwen/qwen35
- Official Blog: qwen.ai/blog
- Qwen Chat: chat.qwen.ai