Qwen3 30B A3B Guide: The Hybrid MoE That Started a Line

Qwen3 30B A3B is easy to misfile. The name reads like a 30B model, the serving cost looks like a 3B model, and the model you download today may not be the one people are benchmarking. The original release from April 2025 is a hybrid: one checkpoint that can either reason step by step or answer directly, switched per request. In July 2025 Qwen replaced that idea with two separate checkpoints, and this article is about the one that came first.

If you are choosing between the three 30B A3B builds, read this for the original, then the Instruct 2507 guide and the Thinking 2507 guide for the split versions.

What the April 2025 release shipped

Qwen released the Qwen3 series on 29 April 2025: two mixture-of-experts models (235B A22B and 30B A3B) and six dense ones from 0.6B to 32B. All of them are open weights under Apache 2.0, which means commercial use, fine-tuning and redistribution without a separate agreement.

The model card gives the shape of the 30B A3B precisely:

• 30.5B total parameters, of which 3.3B are active per token
• 48 layers, grouped-query attention with 32 query heads and 4 key-value heads
• 128 experts, 8 routed per token
• 32,768 tokens of native context, 131,072 with YaRN scaling

Qwen says the family was pretrained on roughly 36 trillion tokens and supports 119 languages and dialects. Its headline claim for this particular model was that it outperforms QwQ-32B, Qwen's earlier reasoning model, which activates about ten times as many parameters. The launch post backs that with a vendor-run comparison chart, so treat it as a direction rather than an independent measurement.

Two modes in one set of weights

The defining feature is the switch. By default the model writes a <think> block of reasoning before its answer. There are two ways to turn that off.

The hard switch is a chat-template argument. With enable_thinking=False the model produces no think block at all:

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

The soft switch works only while thinking is enabled: append /think or /no_think to a user or system message and the model follows the most recent instruction, turn by turn. In soft-switch mode the think block still appears, though it may be empty.

Qwen recommends different sampling per mode. Thinking: temperature 0.6, top-p 0.95, top-k 20. Non-thinking: temperature 0.7, top-p 0.8, top-k 20. The card explicitly warns against greedy decoding in thinking mode, which tends to produce repetition loops. If you serve this model behind an API and leave everything at the provider's defaults, you are probably not running it the way it was tuned.

Why 3.3B active changes the hardware question

Mixture-of-experts splits the cost of a model in two. Total parameters decide how much memory you need to hold the weights; active parameters decide how much compute each token costs. A model is only as cheap as its active count suggests once the weights are already resident.

For this model the gap is large. At two bytes per parameter in BF16, 30.5B parameters is roughly 61 GB of weights before any KV cache, so it does not fit on a single consumer card unquantised. But once it is loaded, each token touches only about 3.3B parameters, so throughput behaves more like a small model than a 30B one. Quantised to 4 bits the weights shrink to something in the mid-teens of gigabytes, which puts it within reach of a single workstation GPU or a high-memory laptop.

The trade is the usual MoE one: you pay memory for capacity you do not compute on every token. If memory is your constraint and compute is not, a dense model of similar quality may suit you better.

The 2507 split, and what it says about this model

In July 2025 Qwen published two replacements: Qwen3-30B-A3B-Instruct-2507 on 30 July and Qwen3-30B-A3B-Thinking-2507 on 31 July. Each does one job. The Instruct build never thinks; the Thinking build always does. Both moved to 262,144 tokens of native context.

Qwen's own comparison is blunt about the original. On its table, the original in non-thinking mode scores 69.1 on MMLU-Pro and 21.6 on AIME25; the Instruct 2507 scores 78.4 and 61.3. In thinking mode the original reaches 70.9 on AIME25 against 85.0 for Thinking 2507. Splitting the modes let each checkpoint specialise, and on Qwen's numbers both specialists beat the generalist at its own game.

That is the strongest argument against defaulting to the original in 2026. It is also worth remembering that newer generations exist: Qwen 3.6 27B, released in April 2026, is a dense model that runs on one GPU. The Qwen family explainer maps where each generation sits.

When the original is still the right download

There are honest reasons to keep it.

• One deployment, both behaviours. If you self-host and can afford only one 30B-class model in memory, the hybrid gives you reasoning and fast answers from a single set of weights. Running both 2507 builds doubles the memory.
• Per-turn control inside a conversation. The soft switch lets an agent think on the planning turn and answer tersely on the tool-result turns without changing model.
• An existing fine-tune. If you have already trained adapters or evaluation baselines on this checkpoint, the upgrade has a real migration cost.

On the hosted side, OpenRouter lists it at $0.12 per million input tokens and $0.50 per million output at the time of writing, with a 40,960-token context. That figure matches the default position-embedding setting in the model's config rather than the 131K YaRN maximum, so check what a given provider actually enables before you plan around long inputs.

A short checklist before you commit

1. Decide whether you need both modes from one deployment. If not, start with the 2507 build that matches your workload.
2. Set the recommended sampling parameters per mode, and never greedy-decode in thinking mode.
3. Enable YaRN only if your prompts actually exceed 32K; static scaling can cost quality on short inputs.
4. Run twenty of your own tasks in both modes and compare cost per completed task. How to benchmark LLMs on your own work has the setup.