Llama 4 Scout Ollama: Settings, Not Working Fixes, Windows App and Alternatives
Running Llama 4 Scout in Ollama on a Windows PC means a download of roughly 60 GB or more, a context setting that decides if it fits in memory, and a handful of errors that repeat. Here are the settings, the fixes, and the alternatives when your hardware says no.
You typed ollama run llama4:scout, waited through a download the size of a small game library, and then got a crawl of one word every few seconds, or an error about memory, or a model that forgets what you said two messages ago. That is the normal first experience with this model, and almost none of it means something is broken. Llama 4 Scout is a 109 billion parameter mixture of experts model, and Ollama has to squeeze it into whatever GPU and RAM you own.
This article walks through the settings that matter, the fixes for the usual failures, how the Ollama Windows app fits in, and what to use when your machine simply cannot carry the model.
💡 Short version: Scout needs far more memory than its "17B active" label suggests. Set a realistic context length, confirm the model is sitting on your GPU with ollama ps, and if the numbers do not work, use a hosted version instead of fighting your hardware.
What Llama 4 Scout Really Is
Meta released Scout as the smaller sibling of Maverick in the Llama 4 family. The numbers people quote are accurate but easy to misread:
Spec
Llama 4 Scout
Total parameters
About 109 billion
Active parameters per token
17 billion
Experts
16
Advertised context window
Up to 10 million tokens
Default Ollama tag
llama4:scout (Q4_K_M quantization)
Typical download size
Roughly 60 to 67 GB depending on the source
The trap is the "17B active" figure. In a mixture of experts model only a slice of the network runs for each token, which makes generation fast when everything is loaded. But all 109 billion parameters still have to live somewhere in memory. So the speed of a 17B model comes with the storage bill of a 109B model.
The 10 million token context is a capability of the architecture, not something a home PC can use. The memory used by the KV cache grows with every token of context, so a huge window can exhaust VRAM even when the weights fit.
Who Should Run It Locally
Local Scout makes sense if you have a workstation GPU or a machine with lots of unified or system memory, you need prompts to stay on your own disk, or you want to experiment with offline workflows. If you have a single 8 GB or 12 GB gaming card, expect partial offloading to the CPU and slow output.
Hardware Reality Check
Third-party guides put the practical minimum for the quantized model at about 24 GB of VRAM, with the full Q4_K_M build wanting something closer to 70 GB of combined memory to run comfortably. Treat those as rough targets, not promises.
Your setup
What to expect
8 to 12 GB GPU, 32 GB RAM
Most layers on CPU, very slow, may fail to load
24 GB GPU, 64 GB RAM
Works with partial offload, usable for short chats
48 GB or more of GPU memory
Mostly or fully on GPU, good speed
Unified memory machine with 64 GB or more
Works well, speed depends on memory bandwidth
You also need disk space. Ollama stores the model files under your user profile by default, and a 60+ GB download on a nearly full system drive is a classic cause of a failed pull.
💡 Quick test: free about 80 GB on the drive that holds your Ollama model folder before you start. That covers the model plus temporary download data.
Install and Run on Windows
Ollama runs natively on Windows 10 and 11. You download the installer from the Ollama website, run it, and the app starts in the system tray. No WSL setup is needed.
Pull and Run the Model
Open PowerShell or Windows Terminal and run:
ollama pull llama4:scout
ollama run llama4:scout
The first command downloads the weights. The second opens an interactive chat. If you only want to confirm the install works before committing to the big download, pull a small model first, then come back to Scout.
Check Where the Model Lives
Right after the model answers, run:
ollama ps
Look at the PROCESSOR column. "100% GPU" is what you want. A split such as "40%/60% CPU/GPU" means part of the model is running on the processor, which explains slow output. This single command diagnoses more complaints than anything else.
The Settings That Matter
Context Length
This is the setting most people get wrong. Ollama chooses a default context size based on your VRAM: under 24 GiB of VRAM it defaults to roughly 4k tokens, and between 24 and 48 GiB it goes to about 32k. A 4k window means long documents get cut and the model appears to "forget" earlier messages.
You have three ways to raise it:
Ollama app slider. Open the app settings and move the context length slider.
Environment variable. Set OLLAMA_CONTEXT_LENGTH to a value such as 16384 before the server starts.
Per session. Inside ollama run, type /set parameter num_ctx 16384.
Every increase costs memory. Doubling the window roughly doubles the KV cache, so raise it in steps and watch ollama ps after each change.
Setting Environment Variables on Windows
Quit Ollama from the system tray.
Open Settings, search for "environment variables", and open the editor for your account.
Add a new user variable, for example OLLAMA_CONTEXT_LENGTH with the value 16384.
Click OK and start Ollama again from the Start menu.
If you set the variable and nothing changes, the usual reason is that Ollama was still running in the tray and never reread the environment.
Settings Worth Knowing
Setting
What it does
When to touch it
OLLAMA_CONTEXT_LENGTH
Default context for all models
Long documents, long chats
OLLAMA_KEEP_ALIVE
How long the model stays loaded
Set higher to avoid slow reloads of a huge model
OLLAMA_MODELS
Folder where models are stored
Move the 60+ GB files to a bigger drive
OLLAMA_FLASH_ATTENTION
Memory-friendlier attention
Try it when context pushes you over the VRAM line
OLLAMA_KV_CACHE_TYPE
Compresses the KV cache (for example q8_0)
Squeeze more context into the same memory
temperature (Modelfile)
Randomness of output
Lower for facts and code, higher for writing
A custom Modelfile lets you save your choices:
FROM llama4:scout
PARAMETER num_ctx 16384
PARAMETER temperature 0.4
Then run ollama create scout-16k -f Modelfile and start it with ollama run scout-16k.
Not Working? Fixes by Symptom
The Download Fails or Stalls
Check free disk space on the models drive. A 60+ GB pull dies quietly when the disk fills up.
Run ollama pull llama4:scout again. Pulls resume instead of restarting.
On a corporate or school network, a proxy or filter can interrupt large downloads. Try a different connection.
Confirm your Ollama version is current. Newer model architectures need a newer runtime, and an old install may not recognize the tag.
"Model Requires More System Memory"
This is Ollama telling you the weights plus the context do not fit. Your options, from easiest to hardest:
Lower num_ctx to 4096 and try again.
Close the browser and other apps that hold RAM and VRAM.
Increase the Windows page file so overflow has room, knowing it will be slow.
Add RAM, or switch to a smaller model.
It Runs But Is Painfully Slow
Run ollama ps. If the model is split across CPU and GPU, that is your answer. Speed drops sharply once layers spill into system RAM, because the CPU path has far lower bandwidth than VRAM. Lower the context length, enable flash attention, or accept that this hardware is better served by a smaller model.
It Forgets the Conversation
That is the 4k default window again. Raise num_ctx as described above and test with a long prompt. If a long document gets silently truncated, the window is still too small for it.
It Uses the CPU Instead of the GPU
Update your graphics driver. Old drivers are a common cause.
Make sure no other program is hogging the GPU memory. Games, video editors and browser tabs with hardware acceleration all take their share.
Restart Ollama from the tray so it redetects the card.
If you have both an integrated and a dedicated GPU, check that Windows is not routing the process to the weaker one.
The App Opens but Nothing Responds
Quit Ollama in the tray, make sure no stray ollama process remains in Task Manager, then start it again. If another program is already using the default port, change it with OLLAMA_HOST, for instance 127.0.0.1:11500.
💡 Rule of thumb: when something misbehaves, change one setting at a time and check ollama ps after each change. Changing three things at once hides the cause.
The Windows App Versus the Terminal
The Ollama Windows app gives you a simple chat window, a model picker and a settings panel with the context slider, plus the background server that other tools talk to. It is the friendliest way to use Scout if you do not like command lines.
Feature
Windows app
Terminal
Chat interface
Yes
Basic text
Context slider
Yes
Use num_ctx or the environment variable
Custom Modelfiles
No
Yes
Scripting and automation
Limited
Full
Diagnostics (ollama ps, logs)
Limited
Full
Most people end up using both: the app for casual chat, the terminal for troubleshooting. Whatever you use, the background server also exposes a local API, so editors and third-party chat front ends can connect to it.
💡 Tip: if you set an environment variable and the app slider shows a different value, the slider wins for chats started inside the app. Pick one place to manage context and stick with it.
When Local Is the Wrong Choice
Be honest about the trade. A big local model gives you privacy and no per-message cost, but you pay with disk space, electricity, setup time and patience. If any of these describe you, skip the local route:
You have less than 24 GB of VRAM and under 64 GB of RAM.
You want fast answers for writing, summaries or code, not a weekend project.
You switch between devices, including a laptop that cannot hold the model.
You do not want to maintain drivers, runtimes and a 60 GB model folder.
The model is the same family either way. Hosting it somewhere else only changes where the 109 billion parameters live.
The Memory Math in One Look
Rough planning works like this: weights (around 60 to 67 GB for the default quantization) plus the KV cache (grows with context) plus overhead for the runtime and the operating system. If that total is bigger than your VRAM, the surplus goes to system RAM, and each step down that ladder costs speed. If the total is bigger than VRAM plus RAM, the model will not load at all.
How to Use Llama 4 Scout Instruct on PicassoIA
If you want Scout without the download, Llama 4 Scout Instruct is available on PicassoIA and runs in the browser. There is no installer, no driver check and no context setting to tune against your VRAM.
Write your prompt in the input box. Be specific about format, length and audience.
Adjust the generation settings shown on the page. A lower temperature (around 0.2 to 0.5) suits facts, code and summaries, while a higher one (0.7 to 1.0) suits brainstorming and creative writing.
Set a sensible maximum length for the reply so you do not wait on text you will not use.
Run the prompt, read the result, and refine. Short follow-up instructions such as "shorter" or "as a table" work well.
Prompt Tips
Give the model a role and a goal: "Act as a technical editor. Rewrite this paragraph for a beginner."
Paste the source text directly, then ask for the output format you want.
For long tasks, split the work into stages: outline first, draft second, polish third.
💡 Handy use: run the same prompt on a hosted Scout and on your local Ollama setup. If the answers match in quality, you know your local settings are right.
Alternatives Worth Trying
If Scout is too heavy or not the right fit, these models on PicassoIA cover most of the same jobs:
For local use, smaller models in Ollama are the realistic route on modest hardware. Anything in the 7B to 14B range loads quickly on a mid-range GPU, handles everyday chat, and avoids every memory problem above. The trade is lower quality on long, complex tasks.
Which One to Pick
Privacy matters most: run a smaller model locally in Ollama.
Quality matters most and your hardware is modest: use a hosted model.
You need reasoning: try Deepseek R1.
You work with images and text together: try Qwen3.7-Plus.
Quick Fix Checklist
Before you give up on a local Scout setup, run through this list:
Free at least 80 GB of disk space.
Update Ollama and your graphics driver.
Quit Ollama from the tray and start it again.
Run ollama ps and read the PROCESSOR column.
Lower num_ctx to 4096, then raise it in steps.
Close apps that use VRAM.
Try OLLAMA_FLASH_ATTENTION and a compressed KV cache.
Compare with a smaller model to confirm the install itself is healthy.
Try It Without the Setup
If you would rather spend your time writing, coding and testing ideas than tuning context lengths, open Llama 4 Scout Instruct on PicassoIA and run your first prompt in a minute. Compare it with Deepseek V3.1 and Llama 4 Maverick Instruct on the same task, and keep the one that gives you the best answers. When you want to go beyond text, PicassoIA also offers image generation, video, speech and music models, and you can browse all of them at picassoia.com/en/all-models.