Skip to content

Qwen

What it is

Qwen is a series of Large Language Models (LLMs) developed by Alibaba Cloud, including general-purpose (Qwen), coding (Qwen-Coder), and vision (Qwen-VL) models. The family features standout agentic variants such as Qwen 3.6-35B-A3B (latest frontier variant), Qwen 3.5-Max-Preview, and Qwen 3.5-Plus, which continue to push the boundaries of reasoning and agentic performance. It remains one of the most capable model families available, particularly strong in coding, mathematics, and complex multi-agent workflows.

What problem it solves

Provides high-performance, open-weight alternatives to proprietary models like GPT-4o. It enables powerful local inference for coding assistants and private reasoning tasks without relying on cloud APIs.

Where it fits in the stack

LLM / Reasoning Engine (Open-weights). It can be used as a backend for local agents or via various inference providers.

Typical use cases

  • Local Coding Assistance: Using Qwen3.6-35B-A3B and Qwen2.5-Coder for IDE completions and agentic refactoring. Qwen 3.5 4B has demonstrated the ability to "vibe code" fully working OS web apps in one go.
  • Agent Swarms: Leveraging the agentic reasoning and thinking preservation introduced in Qwen 3.6 for massive parallel workflows and complex reasoning tasks.
  • Multilingual Applications: Leveraging its strong performance across 29+ languages.
  • Large Context Analysis: Utilizing the 256K context window for deep document processing.
  • Edge Deployment: Running smaller variants (e.g., 0.8B, 1.5B, 3B, 4B) on mobile or low-power devices. The 0.8B model is capable of running on a watch.
  • Hosted agent backends: Using frontier Qwen variants through providers such as NVIDIA NIM or OpenRouter when you want multimodal and tool-calling support without self-hosting the biggest checkpoints.

Hosted inference notes

NVIDIA's March 2026 model card for qwen3.5-122b-a10b and OpenRouter's April 2026 release of qwen3.6-35b-a3b are useful signals for how Qwen is being packaged for production inference:

  • Qwen 3.6-35B-A3B: A 35B parameter MoE model with only 3B active parameters per token. It uses a hybrid sparse mixture-of-experts architecture combining Gated DeltaNet linear attention with standard gated attention layers.
  • Context & Multimodal: These models support a 262K token native context window (extensible to 1M via YaRN) and accept text, image, and video inputs.
  • Agent Readiness: Published deployments explicitly support function calling, structured output, and "integrated thinking mode" with reasoning traces.

Getting started

Installation (via Ollama)

The easiest way to run Qwen locally is through Ollama.

ollama run qwen2.5-coder:7b
# Qwen 3.6 variants arriving shortly to Ollama library

Minimal Python Example (via OpenAI-compatible API)

If running via Ollama, you can use the OpenAI client:

from openai import OpenAI

client = OpenAI(
    base_url='http://localhost:11434/v1',
    api_key='ollama', # required but unused
)

response = client.chat.completions.create(
  model="qwen2.5-coder:7b",
  messages=[
    {"role": "user", "content": "Write a python function to calculate fibonacci numbers."}
  ]
)
print(response.choices[0].message.content)

Strengths

  • State-of-the-Art Coding: Qwen3.6 and Qwen3.5 variants set new bars for coding performance. The Qwen 3.6-35B-A3B variant achieves frontier-style coding and agent performance with only 3B active parameters, making it highly attractive for real deployment.
  • Thinking Preservation: Introduces the ability to retain reasoning context from historical messages to make iterative agentic work more stable and efficient.
  • Efficient Architecture: Qwen 3.6-35B-A3B utilizes roughly 3B active parameters, providing a massive performance-to-compute ratio under an Apache 2.0 license.
  • Native Long Context: Supports up to 256K tokens natively, ideal for large codebases. The tiny 0.8B model has demonstrated the ability to reason over a 100-file repository.
  • Wide Model Range: Scales from tiny edge models (0.8B, 2B, 4B) to massive 72B+ and 122B parameter powerhouses.
  • Multimodal Capabilities: Qwen3.6-35B-A3B is a multimodal causal language model with a vision encoder, supporting image and video input.
  • Community Optimizations: ik_llama.cpp dramatically outperforms mainline for Qwen3.5/3.6 on CPU. Unsloth provides optimized GGUF updates for the series.

Limitations

  • Hardware for Large Models: The 72B and 122B MoE models require significant VRAM (40GB+ even with quantization).
  • Nuance in Western Contexts: Like other non-Western models, it may have different cultural biases or instruction-following nuances compared to Llama or GPT.

When to use it

  • For local development where data privacy is paramount.
  • When you need a top-tier coding model that can be self-hosted.
  • For tasks requiring long-context retrieval or reasoning (e.g., repository-level analysis).

When not to use it

  • If you lack the hardware to run models larger than 7B comfortably.
  • If your workflow is strictly tied to a proprietary ecosystem (e.g., exclusive use of Claude Artifacts).

Licensing and cost

  • Open Source: Yes (Apache 2.0 or Qwen License depending on version)
  • Cost: Free (Self-hosted) / Paid (via providers like Groq, Together AI, or OpenRouter)
  • Self-hostable: Yes

Sources / References

Contribution Metadata

  • Last reviewed: 2026-05-11
  • Confidence: high