Qwen3 14B Guide: One Dense Model, Two Ways to Answer

Before Qwen3, choosing a small open model meant choosing a behaviour. You picked a chat model that answered quickly, or a reasoning model that thought at length, and lived with that for every request. Qwen3 14B does both: the same weights can reason step by step or reply directly, and you choose per request.

That is the main thing to understand about it, and also the main thing to get wrong. A hybrid model used with the wrong settings, or with thinking left on for trivial requests, is slower and more expensive than it needs to be. This guide covers how to use it deliberately.

Where it sits in the family

Qwen released Qwen3 on 29 April 2025: six dense models from 0.6B to 32B, plus two mixture-of-experts models. The 14B is a dense transformer with 14.8B total parameters (13.2B non-embedding), 40 layers and grouped-query attention with 40 query heads and 8 key-value heads. It is published under Apache 2.0.

The family was pretrained on about 36 trillion tokens across 119 languages and dialects, roughly double Qwen2.5's data. Qwen claims Qwen3-14B-Base performs on par with Qwen2.5-32B-Base. Note that this is a base-model comparison; it says something about pretraining efficiency, not directly about the instruct model you will call.

How the thinking switch actually works

There are two levels of control.

• The hard switch is enable_thinking in the chat template, which defaults to true. Set it to false and the model never reasons.
• The soft switch applies when thinking is enabled: adding /think or /no_think to a user or system message toggles it, and the most recent instruction wins in a multi-turn conversation.

Hosted endpoints expose this differently. Some map it to a reasoning parameter, some only honour the soft switch in the prompt. Check which one your provider supports before relying on either.

One detail that breaks home-grown integrations: in multi-turn conversations, the model card says to keep only the final answer in the history and drop the thinking content. Qwen's chat template handles that. If you build prompts yourself, you have to strip it, or the context fills with old reasoning and quality degrades.

Sampling settings matter more than usual

The model card gives separate recommended settings for each mode:

• Thinking — temperature 0.6, top-p 0.95, top-k 20, min-p 0.
• Non-thinking — temperature 0.7, top-p 0.8, top-k 20, min-p 0.

It also says, in capitals, not to use greedy decoding in thinking mode, because it can cause performance degradation and endless repetition. If you see a reasoning trace looping, check temperature before blaming the model. A presence penalty between 0 and 2 can help further, with the caveat that higher values may cause language mixing.

Why OpenRouter says 40,960

OpenRouter lists Qwen3 14B with a 40,960-token context and 16,384 maximum output tokens at the time of writing. That odd number comes straight from the model: native context is 32,768 tokens, and the default max_position_embeddings in the shipped configuration is 40,960. The model card explains the split as 32,768 tokens reserved for output and 8,192 for a typical prompt, so read it as a limit for prompt and output combined, not as a promise of 40K of useful input.

The model card says it extends to 131,072 tokens with YaRN, but the common frameworks implement static scaling, which can hurt short-text quality, and the card advises against enabling it unless you need the length. Assume roughly 32K of usable input unless your provider says it has turned YaRN on.

Thinking tokens and the bill

OpenRouter lists it at $0.12 per million input tokens and $0.24 per million output at the time of writing. Cheap, but thinking mode changes the arithmetic: Qwen recommends allowing 32,768 output tokens for most queries, and 38,912 for competition-level maths and programming problems. A request that would take 200 tokens to answer directly might spend several thousand reasoning first.

That is fine for a maths problem and wasteful for a classification. The practical pattern is to default to non-thinking and enable thinking only for request types where you have measured a gain. The token cost of reasoning models works through how quickly that adds up.

14B, 8B or 32B

The three upper dense sizes share the same design and switch, so the choice is about quality per unit of hardware.

• Pick Qwen3 8B for high-volume, simple tasks, or when you need to fit a single smaller GPU.
• Pick 14B when 8B fails on multi-step reasoning or instruction following but you still want something that runs comfortably on one 48 GB card in 16-bit, or a 24 GB card quantised.
• Pick Qwen3 32B when 14B is close but not reliable, and you have the hardware for roughly twice the weights.

For a sense of what a newer generation offers at a similar footprint, Qwen 3.6 27B was released in April 2026 as a dense model that runs on a single GPU. If you are starting from scratch rather than maintaining a Qwen3 deployment, test that as well.

A sensible rollout

1. Start with thinking off and the non-thinking sampling settings.
2. Build an eval of 100 real requests and score them.
3. Re-run only the failures with thinking on. If most flip to correct, route that request type to thinking mode.
4. Cap output tokens per route so a runaway trace cannot surprise you.

That gives you a cheap fast default with reasoning where it pays, from one set of weights.