Qwen 3.6 27B locally: memory by quantization

#Qwen 3.6 27B: the short answer

To run Qwen 3.6 27B without spilling over to the CPU, plan on a card with 24 GB of VRAM, or a Mac with at least 32 GB of unified memory. A 24 GB Mac can work for a test with a short context and few other applications open. Below that, the model may still load, but part of the computation moves to the CPU.

Commande

ollama run qwen3.6:27b. Le tag existe dans la bibliothèque officielle d'Ollama (relevé du 5 septembre 2026).

File

18 GB on Ollama (checked September 29, 2026); 16.10 GB on the LM Studio page (checked September 5, 2026).

Other quantizations

About 19 GB at Q5, 29 GB at Q8, and 54 GB at FP16, according to the QuelLLM catalog. These are estimates, not official file sizes.

Context

Ollama allocates 4K tokens under 24 GiB of VRAM, 32K from 24 to 48 GiB, and 256K from 48 GiB up. Any increase uses memory on top of the file.

Verification habit

ollama ps: as long as the PROCESSOR column doesn't show 100% GPU, the model is spilling over.

#What you're installing

Qwen 3.6 27B is a dense model from Alibaba, released in April 2026 under the Apache 2.0 license. “Dense” means that all 27 billion of its parameters take part in computing each token. The Ollama page lists vision, tools, and thinking capabilities, and an advertised maximum context of 256k tokens. Full specifications are on the /modele/qwen36-27b page: this guide covers only memory and installation.

This model is no longer the newest in its size: Qwen 3.8 27B succeeded it in August 2026. 3.6 is still worthwhile if your prompts are already tuned for it, or if a tool you use references it in its documentation. For a first install, also compare it with our guide to Qwen 3.8 27B, which has its own tags and settings.

#Qwen 3.6 27B memory by quantization

Quantization reduces the precision of the weights to reduce their size. The stronger it is, the smaller the file, with a risk of quality loss. The table deliberately separates published sizes from catalog estimates, because these figures don't carry the same weight.

The “Size + 15 to 25%” column applies a rule of thumb for caution, the same one used in our guide to Qwen 3.8 27B. It's meant to account for the context cache and engine overhead. It's a calculation, not a measurement: actual usage depends on the chosen context, the engine, and whatever else is running alongside it.

Why 18 GB and 16.10 GB?

Ollama and LM Studio don't necessarily distribute the same package. Our data doesn't let us attribute the difference to a specific cause: go by the figure shown by the tool you're using.

Why does the catalog list 16 GB in Q4?

It's an estimate based on the parameter count. It gives a ballpark figure and comes in 2 GB below the file size listed by Ollama. For this tag, the Ollama figure takes precedence.

A point of comparison

For Qwen 3.8 27B, a model of the same size, Ollama listed 18 GB in Q4_K_M, 30 GB in Q8, and 56 GB in BF16 (checked August 18, 2026). These values apply to a different model: they indicate a trend, but don't replace data for 3.6.

If several variants fit in your memory, start with Q4_K_M: it's the recommended setting on QuelLLM. Only move up if your own prompts reveal a flaw that quantization explains. Our guide to choosing a quantization covers this tradeoff in general.

#The headroom to keep for context

The file size is only the starting point. The engine also reserves memory for the context window, meaning the conversation, documents, and the model's reasoning. The advertised 256K-token context is a maximum, not what will actually be allocated.

Less than 24 GiB of VRAM

Ollama starts at 4k tokens. That's enough for a short exchange, but not for a long document.

From 24 to 48 GiB

32k tokens by default: the range of a RTX 4090.

48 GiB and up

256k tokens by default. This allocation takes up a lot of space: reduce it if you don't need it.

Agents, code, web search

Ollama recommends at least 64,000 tokens for these uses. On 24 GB, that value isn't guaranteed: test it with ollama ps.

The risk is twofold. A window that is too short forces the engine to leave out part of the input, and the model then responds based on incomplete text. A window that is too long can make the model spill over onto the CPU. Since we have no measurements for this model, this guide gives no GB figure per token range: the right approach is to increase it in steps.

1. 01
   Start from the default value

   Run the model with a typical request, then check ollama ps. Note the memory used, the context, and the GPU/CPU split.
2. 02
   Increase in steps

   For example, go from 8K to 16K, then 32K and 64K tokens. Reload the model after each change: the context is set at load time.
3. 03
   Stop before you overflow

   As soon as PROCESSOR shows a split between CPU and GPU, go back to the previous step. Leave some headroom if you also send images, since processing them uses memory too.

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

curl http://localhost:11434/api/generate -d '{
  "model": "qwen3.6:27b",
  "prompt": "Résume ce texte en trois points : ...",
  "stream": false,
  "options": { "num_ctx": 32768 }
}'

If Ollama is already running as a service, the variable must be set in the service configuration, because a command run in a terminal doesn't change it. The Ollama app also offers a context slider in its settings.

#Is your machine enough?

The memory table translates fairly simply into hardware. The guidelines below apply to the default 18 GB tag. They follow from this size calculation and don't describe tests: no speed is claimed.

RTX 3060 12 GB, RTX 4070 12 GB

The file exceeds your VRAM: the model will run partly on the CPU. Choose a smaller model instead, around 14B in Q4 (≈ 9 GB).

RTX 4080 16 GB

The 18 GB file still exceeds VRAM, as does the 16.10 GB of LM Studio once context is added. It remains usable with a GPU/CPU split, though not comfortably.

RTX 4090 24 GB

The first sensible tier: it leaves about 6 GB above the file for context and engine overhead.

Mac M4 Pro 24 GB

Possible for a test: the system and applications share the same memory as the model. Keep the context short and close heavy applications.

Mac with 32 to 48 GB

The sensible choice for daily use, with real headroom for context.

Without a graphics card

The QuelLLM catalog estimates the required RAM at 28 GB. The model can run, but expect much slower generation than on a GPU.

You also need disk space: at least the size of the file, with some margin if you keep multiple variants. For your specific configuration, the QuelLLM configurator gives a first estimate, which you should confirm with ollama ps.

#Install Qwen 3.6 27B with Ollama

Ollama is a good fit if you want a local service accessible via API, on http://localhost:11434, optionally with Open WebUI as the interface.

1. 01
   Install or update Ollama

   On Linux, use the official script below. On macOS and Windows, download the application from ollama.com. An older version may not recognize the model: updating is the first thing to do if you get an error.
2. 02
   Download the model

   ollama pull qwen3.6:27b récupère le fichier. Si la connexion coupe, relancez la commande.
3. 03
   Check what you received

   ollama show qwen3.6:27b affiche notamment la quantification, le contexte maximal et les capacités du paquet. C'est le moyen le plus sûr de savoir à quelle variante correspond le tag par défaut.
4. 04
   Start a first exchange

   The first response arrives after the weights have loaded into memory. While the model is loaded, read ollama ps in a second terminal.

# Linux : installer ou mettre à jour Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Télécharger le modèle
ollama pull qwen3.6:27b

# Vérifier la quantification, le contexte et les capacités
ollama show qwen3.6:27b

# Discuter avec le modèle
ollama run qwen3.6:27b

# Dans un second terminal : mémoire et répartition GPU/CPU
ollama ps

On Apple Silicon Macs, the library also offers qwen3.6:27b-mlx (checked September 5, 2026), intended for the MLX engine. Its size may differ: check the tags page before choosing.

#Install with LM Studio

LM Studio is a good fit if you prefer a graphical interface, with visible settings for context and the number of layers offloaded to the GPU. Its official page uses the identifier qwen/qwen3.6-27b, lists 16.10 GB, and confirms support for tool use, vision, and reasoning (checked September 5, 2026).

1. 01
   Search for the model

   In the model search, enter qwen/qwen3.6-27b. LM Studio lists the available variants and highlights the one it considers suited to your hardware.
2. 02
   Choose the format

   GGUF on PC, MLX or GGUF on Mac. Compare the displayed size with your memory, keeping the margin from the previous table.
3. 03
   Configure loading

   Before loading the model, adjust the context length and GPU offload. Start with a modest value, then increase it in steps, as with Ollama.
4. 04
   Monitor memory

   Run a test with a real prompt and check the memory used. If the system slows down, reduce the context before changing the quantization.

# Variante GGUF (--mlx sur Mac pour le format MLX)
lms get qwen/qwen3.6-27b --gguf

You can then expose the model to other applications through the LM Studio local server: our guide to its API covers this step in detail. To choose between the two tools, see our Ollama and LM Studio comparison.

#Qwen 3.6 27B or 35B-A3B: the choice in principle

The two models belong to the same family but aren't built the same way. 27B is dense. 35B-A3B is a Mixture-of-Experts (MoE) model: it has 35 billion parameters, of which about 3 billion are active per token. The comparison below is based on this architecture and on published sizes, without benchmarks.

Memory

Advantage 27B: 16.10 GB versus 20.40 GB on the LM Studio pages (checked September 5, 2026). An MoE model must keep all its experts in memory, even if few of them work on each token.

Per-token computation

Advantage 35B-A3B in principle: fewer active parameters mean less computation per token, provided the whole model fits in memory. The actual difference depends on your machine.

If you're right at the 24 GB limit

27B leaves more room for context. 35B-A3B leaves less, which can matter with long documents.

If you have 32 GB or more

Both fit. Test them on your own tasks: the 35B-A3B if responsiveness is the priority, the 27B if vision matters (its capabilities are confirmed by our data).

License

Apache 2.0 for both: this criterion doesn't separate them.

#Troubleshooting

"model not found" during download

Check the spelling: qwen3.6:27b, with a period. If a quantization suffix is rejected, it may not exist: check the tags page, then update Ollama if the tag is listed there.

Very slow generation

Check ollama ps. A CPU/GPU split indicates overflow: reduce the context, close applications using the GPU, or switch to a smaller model.

The beginning of the document seems to be ignored

The context is probably too short, especially with the default 4K tokens under 24 GiB. Increase num_ctx or split the document.

The OLLAMA_CONTEXT_LENGTH setting has no effect

The variable must apply to the server, not the client. Restart the service after changing it, then reload the model.

Out-of-memory error on load

Lower the context first. If the error persists, check free memory and the number of loaded models: ollama stop qwen3.6:27b frees this model.

Sizes that differ from this guide

Packages can change. Rely on the official page at the time and on what your tool displays, then apply the same margin.

#Go further

These pages complement the guide without repeating its figures: the model page for specifications, and related guides for other models and general concepts.