Qwen 3 locally: complete test and real benchmarks

◆ Local AI — Your private ChatGPT, free, on your own machine, in an hour · $24 · or all kits $49 →

In brief

Qwen 3 comes in dense versions (8B, 14B, 32B) and an MoE version (30B-A3B, 3B active parameters out of 30). · On RTX 4090, expect about 95 tok/s for 8B, 78 tok/s for 14B, and 110 tok/s for 30B-A3B—the MoE is faster than the dense 8B. · On RTX 3060 12 GB, the 14B (about 32 tok/s) remains the best compromise; the 30B-A3B does not fit in this card's VRAM. · On a Mac M4 Pro, expect 38, 24, and 45 tok/s, respectively.

#The Qwen 3 lineup: dense vs. MoE

Alibaba released Qwen 3 in two distinct families. Dense models (Qwen3-1.7B, 4B, 8B, 14B, 32B) follow the standard architecture: all parameters are used for every token. The MoE family (Qwen3-30B-A3B, Qwen3-235B-A22B) activates only a fraction of the experts for each token—hence the A3B suffix = 3 billion active parameters out of 30 total.

In practical terms, for a local user, two things change: the required VRAM corresponds to the total parameters (the entire model must fit in memory), but speed corresponds to the active parameters. A 30B-A3B MoE fits in 19 GB in Q4 like a dense 32B, but generates tokens at the speed of a 3B. That’s what makes it so interesting.

The three test variants

We focus on Qwen3-8B (dense), Qwen3-14B (dense), and Qwen3-30B-A3B (MoE). These are the three sizes that most 12–24 GB configurations can run comfortably.

#Tested setup and hardware

The Local AI Kit

Your private ChatGPT, free, on your own machine in an hour — LM Studio, Ollama, Open WebUI, your documents, no cloud.

Lifetime online access
PDF + files
Lifetime updates

The Local AI Kit →or all kits for life — $49 →

Three machines cover roughly the full spectrum of local setups in 2026:

RTX 4090 24 GB: Windows 11 tower, Ryzen 9 7950X, 64 GB DDR5. High-end consumer hardware.
RTX 3060 12 GB: Linux tower running Ubuntu 24.04, Ryzen 5 5600X, 32 GB DDR4. The most popular budget configuration.
MacBook Pro M4 Pro 48 GB: macOS 15, 16-core GPU, unified memory. Representative of high-end laptops Apple.

Identical backend everywhere: Ollama (daemon on localhost:11434). All models in Q4_K_M by default, 8192-token context, standardized prompt of about 500 tokens. Speed measured over 10 generations averaging 500 tokens.

Retrieve the three models

ollama pull qwen3:8b
ollama pull qwen3:14b
ollama pull qwen3:30b-a3b

→

Measure on your system

To reproduce this test, add --verbose to ollama run. Ollama displays the eval rate (generation tokens/sec) and prompt eval rate (prefill speed) at the end of each response.

#Qwen3-8B: the entry-level dense model

Dense model with 8 billion parameters, ~5 GB in Q4_K_M. Designed to run everywhere: 8 GB of VRAM is more than enough. It is the natural candidate when you discover Qwen 3 on a RTX 3060 or laptop.

RTX 4090: ~95 tokens/sec. The card is largely underutilized — all available VRAM is underused at this level.
RTX 3060 12 GB: ~52 tokens/sec. Very comfortable, well above the fluidity threshold (20 t/s).
M4 Pro 48 GB: ~38 tokens/sec. Slower than RTX cards but highly usable, with zero fan noise.

In terms of quality, Qwen3-8B handles everyday tasks very well: summarization, rewriting, simple code, and Q&A. It is clearly outclassed once you get into multi-step reasoning or nontrivial code—that's where you can tell it's an 8B model. Its French is decent but not excellent, with some anglicisms and awkward phrasing.

#Qwen3-14B: the dense sweet spot

Dense 14B, about 9 GB in Q4_K_M. Easily fits on 12 GB of VRAM with a reasonable context. This is the model we recommend to 90% of people with a RTX 3060 12 GB, 4070, or equivalent.

RTX 4090: ~78 tokens/sec. Excellent speed—we’re still well below the GPU ceiling.
RTX 3060 12 GB: ~32 tokens/sec. Comfortable for chat use; 8k context works without offloading.
M4 Pro 48 GB: ~24 tokens/sec. At the edge of comfortable but usable every day.

The quality jump over the 8B is clear. Correct Python code for 50-80-line functions, stronger reasoning, and noticeably better French (natural phrasing, few gender errors). It is the first model in the lineup that genuinely resembles a low-end cloud assistant in terms of consistency.

Default context Ollama

Ollama defaults to a 2048-token context limit, which is ridiculous for Qwen 3 (which natively supports 32k). Force num_ctx through a Modelfile or with /set parameter num_ctx 8192 in the session; otherwise, you will cut off long conversations.

#Qwen3-30B-A3B: the MoE that changes everything

This is where it gets interesting. The 30B-A3B weighs about 19 GB in Q4_K_M (= you need 24 GB of VRAM to run it comfortably with context), but generates using only 3B active parameters. Result: 30B quality, 3B speed.

RTX 4090: ~110 tokens/sec. Yes, faster than the dense 8B model. That's the magic of MoE: less computation per token.
RTX 3060 12 GB: Doesn’t fit in VRAM. Partial CPU offload ≈ 8 t/s — usable for testing but frustrating.
M4 Pro 48 GB: ~45 tokens/sec. Unified memory shines here: everything fits in RAM, and the MoE offsets the limited FLOPS.

The quality is very close to a dense Qwen 32B. The 30B-A3B is particularly strong at coding (in the same league as a Qwen3-Coder 30B-A3B on intermediate Python tasks) and structured reasoning. For written French, it is genuinely good—roughly at the level of a Mistral Small 24B, for reference.

→

The surprise verdict

If you have 24 GB of VRAM or a Mac with 32+ GB of unified memory, Qwen3-30B-A3B remains an excellent general-purpose MoE: better quality AND better speed than the dense 14B, with no tradeoff. Since this test, the next generation has taken over in this range—Qwen 3.8 27B (the “Copilot-like” option of 2026, with 262k context) and Qwen 3.6 35B-A3B, the safe choice for 24–32 GB configurations. The 30B-A3B measured here retains all of its MoE DNA.

#Thinking mode: useful or not?

Qwen 3 introduces a toggleable thinking mode. When enabled, the model first generates a <think>...</think> block with its reasoning before producing the final answer. It is inspired by DeepSeek R1 but is more modular: you can enable or disable it on the fly.

Enable or disable thinking

# Dans la session Ollama
/set parameter think true
/set parameter think false

# Ou via prompt direct (Qwen 3 reconnaît les balises)
>>> /think Combien fait 17 × 23 ?
>>> /no_think Salut, ça va ?

What we observe from 50 tested prompts:

Simple tasks (chat, paraphrasing): Thinking mode = wasted time. It adds 2 to 5 seconds of latency with zero quality gain. Disable it.
Math, logic, multi-step reasoning: Net accuracy gain: +20 to +30% correct answers on GSM8K-style problems. Keep it enabled.
Complex code (refactoring, debugging): Moderate gain (+10%) in quality, but doubled latency. Test it based on your patience.
Creative generation (long-form text, dialogue): Often counterproductive: the model overthinks and loses spontaneity.

Token cost

A typical <think> block adds 500 to 2000 additional tokens before the response. On a 30B-A3B at 100 t/s, this is invisible. On a 14B at 30 t/s, it adds 10 to 60 seconds of latency— weigh your priorities.

#French quality vs. Llama 4

We compared the French outputs of Qwen3-14B and Qwen3-30B-A3B with Llama 4 Scout (the equivalent Meta model, released around the same time) across three exercises: writing a professional email, explaining a technical concept in plain language, and analyzing a short literary text.

Qwen3-14B: Correct French, natural phrasing, but a few persistent gender errors (“la problème”). Vocabulary is somewhat flat.
Qwen3-30B-A3B: Very good written French. No structural errors, varied vocabulary, and register appropriate to the context. Slightly behind Mistral Small 24B on highly idiomatic nuances.
Llama 4 Scout: Flawless French grammatically, but with a more “translated from English” tone than Qwen. Personal preferences.

Pragmatic verdict: for professional French, Qwen3-30B-A3B and Llama 4 Scout are neck and neck. The choice comes down to other criteria (speed, size, license). For literary or highly idiomatic French, Mistral Small remains a notch ahead.

#Which Qwen 3 should you choose for your setup?

6-8 GB of VRAM (RTX 3050, 3060 Ti, 4060): Qwen3-8B in Q4_K_M. The 14B runs in Q3, but quality deteriorates — stick with the clean 8B.
12 GB of VRAM (RTX 3060 12, 4070, 5070): Qwen3-14B in Q4_K_M is the undisputed sweet spot. Comfortable 16k context.
16 GB (RTX 4080, 5070 Ti, 5080): Qwen3-14B in Q5 or Q6 for comfort, or a 30B-A3B in Q3_K_M if you want to test the MoE.
24 GB (RTX 3090, 4090, RX 7900 XTX): Qwen3-30B-A3B in Q4_K_M. No debate—it’s the best general-purpose local model available today.
Mac M-series with 32-48 GB unified memory: Qwen3-30B-A3B in Q4 or Q5. The MoE is particularly well suited to unified memory Apple.
Mac Studio Ultra 128 GB+: Jump straight to Qwen3-235B-A22B (~120 GB in Q4) if you want the top of the range. Otherwise, the 30B-A3B in Q8 remains excellent.

→

And what about thinking mode?

By default, leave it disabled. Enable it on demand when tackling a math, logic, or complex reasoning problem. For everyday chat, it only adds latency. The same applies to the next generation: Qwen 3.8 27B tends to overthink with its default reasoning setting — switch it to low for everyday conversations.

#Go further

If Qwen 3 is your starting point, know that the next generation has since taken over in our rankings — Qwen 3.5 9B has become the default choice for 8 GB configurations (256k context, vision), and Qwen 3.8 27B for 24 GB configurations; the settings below remain valid unchanged for these models. A few ways to go further:

Choose your quantization: To understand exactly what you give up between Q4_K_M and Q5_K_M on Qwen 3, and when moving up in quantization makes sense.
Customize with a Modelfile: To set the thinking mode, context, and a default FR system prompt on your Qwen 3 instance—instead of configuring them again each session.
Open WebUI with Ollama: To move from the terminal to a real chat interface with history, Markdown, and attachments without giving up local operation.

Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.

QuelLLM KitsThe reference guide by use case

Local AI Local Copilot Local Agents Local RAG AI at Work Uncensored AI Mac Image AI

All kits for life — $49

Logical next step

Getting started · 9 min

Qwen 3 locally: full test and benchmarks real