On April 29, 2025, Alibaba's Qwen team released Qwen3, a family of eight open-weight models that range from a 0.6B dense model to a 235B mixture-of-experts (MoE) model, all under the Apache 2.0 license. The headline feature was hybrid reasoning: the same Qwen3 model can think step by step for hard problems or answer directly for simple ones, and the caller can switch modes per request. That combination of permissive licensing, broad size coverage and controllable reasoning made Qwen3 one of the most practical open model families for developers who want to run models locally or fine-tune them.

Key Facts#

  • Release date: April 29, 2025, announced on the Qwen blog and GitHub, with weights on Hugging Face and ModelScope.
  • Dense models: 0.6B, 1.7B, 4B, 8B, 14B and 32B parameters.
  • MoE models: Qwen3-30B-A3B (about 3B active parameters) and the flagship Qwen3-235B-A22B (about 22B active).
  • License: Apache 2.0 for all of the open-weight models, which allows commercial use, modification and redistribution.
  • Hybrid thinking: a thinking mode for math, coding and logic, and a non-thinking mode for fast chat, selectable per request.
  • Languages: support for more than 100 languages and dialects, according to the Qwen team.
  • Follow-up: the Qwen3-2507 updates in July and August 2025 split the models into separate Instruct and Thinking variants, starting with a 235B Instruct model that supports 256K-token context.

What Happened#

The Qwen team described Qwen3 as its most capable release to date, built on lessons from its QwQ reasoning model and the Qwen2.5 series. According to the team's own benchmarks, Qwen3 surpassed QwQ in thinking mode and Qwen2.5 instruct models in non-thinking mode on mathematics, code generation and commonsense reasoning. The team also highlighted better alignment with human preferences for writing and multi-turn dialogue, and stronger agent skills, including tool use in both modes. As always, these are vendor claims.

Developers control the reasoning mode in two ways. In code, passing enable_thinking=False to the chat template makes the template insert an empty think block, so the model skips its step-by-step reasoning. In conversation, users can add the soft switches /think or /no_think to a system or user message, and the model follows the latest instruction in a multi-turn chat.

A Hugging Face analysis of the Qwen3 chat template, published the next day, pointed out several details that matter for agents. The template keeps the model's reasoning for assistant turns after the most recent real user message, so a multi-step tool-calling plan stays visible to the model. It strips older reasoning to save context. It also serializes tool-call arguments only when they are not already strings, which avoids double-escaped JSON, and it does not inject a default system prompt.

The Qwen repository documents support across the usual runtimes: llama.cpp, Ollama, LM Studio and MLX for local use, plus vLLM, SGLang and Hugging Face TGI for server deployments.

Background#

Qwen had become a common base for other open models well before Qwen3. When DeepSeek released its R1 distilled models in January 2025, four of the six were fine-tuned from Qwen2.5 models, which were also Apache 2.0 licensed. The Qwen team released its first MoE model, Qwen1.5-MoE-A2.7B, in March 2024, and the QwQ series explored dedicated reasoning models.

Qwen3 arrived only weeks after Meta's Llama 4, which used a custom license with attribution and naming rules. Qwen3 offered a simpler choice: standard Apache 2.0 terms and models small enough for laptops and edge devices as well as large ones for servers.

Why It Matters for Developers#

Qwen3 lets you trade latency for reasoning inside one deployment. Before hybrid models, teams often ran a fast chat model and a separate reasoning model and routed between them. With Qwen3, you can keep one model loaded and decide per request whether it should think. That simplifies local and edge deployments where GPU memory is scarce.

The small MoE model is especially interesting. Qwen3-30B-A3B activates only about 3B parameters per token, so it generates quickly, although all 30B parameters still have to fit in memory. On a workstation with enough RAM or VRAM, it gives much better quality than a 3B dense model at a similar per-token cost.

For .NET developers, Qwen3 works well with Ollama and OllamaSharp, which implements IChatClient from Microsoft.Extensions.AI. The same abstraction handles tool calling, so you can test agent workflows locally before moving to a hosted model:

C#
using System.ComponentModel;
using Microsoft.Extensions.AI;
using OllamaSharp;

IChatClient client = new ChatClientBuilder(
        new OllamaApiClient(new Uri("http://localhost:11434"), "qwen3:8b"))
    .UseFunctionInvocation()
    .Build();

var options = new ChatOptions { Tools = [AIFunctionFactory.Create(GetOrderStatus)] };

// "/no_think" is Qwen3's soft switch to skip the reasoning phase for simple requests.
var response = await client.GetResponseAsync("Where is order 4711? /no_think", options);
Console.WriteLine(response.Text);

[Description("Gets the shipping status of an order.")]
static string GetOrderStatus(int orderId) => orderId == 4711 ? "Shipped" : "Unknown";

Two practical cautions apply. First, runtime tags do not always match Qwen's own names. The Qwen repository warns that some Ollama tags point to newer 2507 variants, so check the tag list before you pin a model in configuration. Second, the soft switches and enable_thinking apply to the original hybrid models. The later 2507 Instruct models never think, and the Thinking models always do. Treat the mode as part of your model selection, and cover both paths in your function-calling tests.

What's Next#

The Qwen team kept a fast release cadence. Between July 21 and August 8, 2025, it shipped the Qwen3-2507 models, which separated instruction and thinking variants and extended context, first to 256K tokens and later to documented support for inputs of up to 1 million tokens. In September 2025, it released Qwen3-Next-80B-A3B, an ultra-sparse MoE model, followed by Qwen3.5 in February 2026 and further 3.x releases during 2026.

The bet on broad coverage and Apache 2.0 licensing paid off in adoption. Hugging Face's summer 2026 report called Qwen the community's base model, with more than 151,000 derivative models on the Hub and about 39.6 million GGUF downloads per month. If you are choosing an open model family for local AI in .NET, Qwen3 and its successors are a sensible default to benchmark first, alongside the patterns in our guide to AI agent architecture.

Sources#