How to use Green Zone

The information you need to enter

Green Zone calculates a possible configuration from the values you provide. It does not scan your computer, identify its graphics card or measure free memory. You can plan for your current computer or equipment you are considering buying.

Start with two separate capacities. Video Random Access Memory (VRAM) is memory on the Graphics Processing Unit (GPU). Random Access Memory (RAM) is system memory used by the operating system, applications and, when enabled, part of the model. Enter each total in its own field: for example, 8 GiB VRAM and 32 GiB RAM. GiB and GB differ slightly; our fields use GiB.

The RAM field is always visible. Turn on “Allow part of the model to use system RAM” to include it in model placement. You can then set a system reserve and an assumed GPU weight share. Leave room for the operating system and other applications. With this option off, the plan keeps the model entirely on the GPU.

Do not add the capacities together: 8 plus 32 does not give you a 40 GiB graphics card. The calculator checks the two memory pools separately. On a unified-memory computer, do not enter the same shared capacity twice. Our approximate GPU/RAM split is not a separately validated plan for these machines.

Where to start on your first visit

First choose the kind of work: conversation, coding, document research or actions performed by an AI agent. Workload buttons provide an initial allocation. They do not install an application or establish that every model will perform your task well.

If you have chosen a model, start with the first tool, the memory check. If you do not know model names yet, start with the catalog in chapter two. A model may already be selected in the calculator; that is an interface starting point, not a model detected on your computer.

For your first calculation, choose a model and file, enter memory, select a workload and press Balance. You can leave advanced settings alone for now. Then check two results: whether memory is sufficient, and whether there is room for your materials and answer. The following chapters explain each tool separately.

01 · Check whether your chosen model fits

When to use it

You have downloaded a model or want to check a particular choice before downloading it. This tool estimates room for both the model itself and the work you want it to do.

What to do

Select the model, then choose its file under weight quantization. These are versions of the same model that store numbers differently and have different sizes. Enter VRAM and RAM, enabling RAM use if needed. Select a workload and press “Balance this configuration”.

What the result means

Read the memory status and context-budget status first. Context is the working space for text received and generated in a request. A small model file does not mean unlimited documents will fit. Model weights need memory, and processing the conversation needs additional space.

The allocation diagram shows video-memory use. If some of the model is placed in RAM, a separate system-memory estimate appears. The graph below shows how estimated video-memory needs change with context length. It is not a speed graph.

What to do next

The green zone means the estimate fits the selected assumptions. Also check that the text budget is balanced. “Estimated context limit” is a calculated boundary, not a target you must reach. If space is tight, try less text, a smaller file or permitted RAM use. Recalculate and compare.

02 · Find a model without knowing its name

Narrow the choice by memory first

You do not need to know any model names. Enter VRAM and the desired context at the top of the catalog. For a first exploration, you could try 4,096 tokens and later increase it for your documents. This is an example starting point, not a required minimum. To include RAM, configure and enable it in the memory calculator first; settings carry between tools.

How to browse

Leave search blank and keep all families and precision classes available. Enable the calculated-fit filter and sort by file size to begin with less demanding options. Name search becomes useful later if you have encountered a specific model such as Qwen. You do not need to choose a family first.

Open a candidate’s detail page. Review its files, parameter count, context limit and links to the publisher. Size helps estimate requirements but does not establish suitability. For coding, look for information about code support; for conversation, check instruction following and your language. Consult the publisher’s description: the catalog does not test every model’s answer quality.

Continue with your actual task

Choose a file and carry it into the memory calculator. Check your own workload rather than relying on a catalog row alone. If nothing appears, remove family and precision restrictions, request less context or review the budget. No result means our search found no candidate under those conditions, not that the computer cannot run any local model.

03 · Find a configuration for your task

When automatic selection helps

You have selected a model and roughly know the space your work needs. Automatic selection compares available combinations for that model instead of making you try each file manually. It does not choose another model; that is the catalog’s job.

Getting started

Fill in the materials and answer budget in the context tool. Then select a priority: a balanced configuration, more context or higher weight precision. Allow RAM use if appropriate. Cache precision and minimum weight precision can stay at their current values until you have read their explanations.

How it differs from Balance

Automatic selection preserves the budget allocations you entered and looks for a configuration that holds them. Balance can redistribute unlocked allocations. Enable the context lock if the entire requested window must be retained; otherwise a smaller window may be offered, but it must still hold the entered budget.

Compare the proposed file, text capacity, memory requirement and RAM use. Apply a suitable option and inspect it in the main calculator. Higher precision is not a measured improvement in answer quality. If there are no options, reconsider specific requirements rather than increasing every value at once.

04 · Make room for documents, conversation and answers

What this tool counts

Plan how much text the model should read and how much it may write. The unit is a token, a small text fragment, not a fixed number of words. Follow the link for a full explanation if the concept is new.

Build the budget

Start with the current request and documents you will actually supply. Add relevant history and instructions, then leave an answer reserve and, for an agent, space for later steps. Sliders make a plan; they do not upload files or change a running model.

Allocated context is the total capacity. Total Tokens adds all allocations, including the future answer. Remaining space has not been assigned yet. With a window of 8,192 and allocations of 7,206, 986 tokens remain. The answer reserve is already included: 986 is not the whole answer allowance.

When you move a slider

Linked balancing first uses spare space, then expands context when memory permits, then reduces other unlocked allocations. Watch the totals as well as the field you changed. Lock anything that must not be reduced automatically. A balanced text budget still needs the memory check in the first calculator.

05 · Compare files of the same model

Why compare?

You are choosing between a smaller file and higher numerical precision. This tool shows the selected model’s files under the same memory and context settings, so you can change one factor and see its effect.

Q4, Q5, Q6 and Q8 describe roughly 4-, 5-, 6- and 8-bit quantization classes. FP16 and BF16 are different 16-bit formats. You do not need to memorize the labels: look at the actual file size and total estimated memory. A smaller file often leaves more document space but can change answer quality.

Make a choice

Compare memory needs and the estimated context limit. Use the arrow button to select a file and return to your task. Pin a configuration if you want to compare it later. This table measures neither speed nor answer accuracy; test a memory-fitting option on your work. Weight quantization and cache precision compress different data.

06 · Understand the terms and results

This card opens explanations rather than another calculation. Start with tokens, weights and context: they connect model size to the amount of work. Technical terms in this guide are linked to the glossary, which provides a return to the previous page.

Stories illustrate why someone might use a local model and how agents use tools. You can read them independently of the calculators. Method & Data is for exploring sources and calculation assumptions. You do not need to study formulas before your first selection.

Compare different models or computers

Choose a model and its quantization in the selector at the top of this page, then press Add to comparison. You can add up to four options. Each saved card also has model and quantization menus, so you can replace any option without leaving the comparison. The file menu always belongs to the model selected in that card.

Open Memory settings for this option to change its VRAM, RAM or placement. Changes to a card are saved automatically and update the common graph; the other cards and your main calculator stay unchanged. Open in memory calculator is a separate button for returning to the full calculator.

Read the graph together

The slider sets the same context window for every saved option. Each colour belongs to one option: the solid line shows its estimated memory including reserves, and the dashed line shows the capacity you entered. If the solid line is above the dashed line, that option needs more memory at this context. A curve ends at the model’s own context limit. The table below shows the numbers without rounding them to K.

Video memory and system memory

Choose video memory (VRAM) or system memory (RAM) above the graph. RAM appears when you have allowed some of the model to run there; the two capacities are not added together. Other settings, such as cache precision and simultaneous requests, remain as saved for each option. This is a memory comparison, not a test of speed or answer quality. Open any saved option in the calculator to change it, or remove it to make room for another.

What Balance this configuration changes

Balance builds a starting plan for the selected workload. It tries the current file first, then smaller available variants when needed. It can change context, permitted cache format, working-memory allowances, reserves and token allocations. These changes are inside the website; they do not reconfigure your installed model.

With RAM use disabled, selection considers full GPU placement. When enabled, it may split the model between GPU and RAM. Afterwards, review the file, context, GPU share and both results: memory and text budget.

Locks protect allocations from automatic reduction. If protected values do not fit, check them and manually reduce only what is acceptable, or increase available space. Pressing Balance again may make no change when a suitable plan is already selected.

When you choose another model in the memory calculator or open it from the catalog, Green Zone creates a fresh example task. It keeps your entered VRAM and RAM, replaces the previous token allocations and locks, and starts with one request at a time. Chat starts at up to 8,192 tokens, coding and agent work at up to 16,384, and research at up to 32,768; less is used when memory or the model requires it. Save or pin your earlier setup if you want to return to it. Saved comparison cards keep their own settings.

If Balance cannot find a plan, read the next steps inside the memory result. They distinguish a model’s context limit from insufficient VRAM or RAM. Memory suggestions name the selected file and context. One option places the whole model on the GPU; another uses your entered VRAM together with RAM. RAM estimates include your operating-system reserve. You can explicitly try the suggested capacities in the calculator. Nothing measures or changes your computer. A larger memory option adds headroom, but does not predict speed or runtime support.

Manual Balance and workload buttons preserve your locked token allocations. If those belong to an old task, “Start a new example task” replaces them and their locks while retaining your hardware values. A locked context is only shortened when you allow it.

What goes in each budget field?

System instructions define the assistant’s role and rules. Tool descriptions explain available actions and their arguments, not all the tool’s program code.

Loaded skill instructions are guides actually read by the model. An installed skill folder does not automatically consume tokens: count the instructions and supporting material included in the request.

Files and documents are supplied text, excerpts or code. History contains previous messages included now. Current task is your new request. Chat-template allowance leaves room for message formatting added by the application.

Answer reserve makes room for generated text, including reasoning when present. Future agent steps allow for new tool results and continuation. Do not count the same text twice. A separate sub-agent may have its own context and return only a summary.

Estimating when you do not have exact counts

Start with a small sample of your work. For illustration, a brief skill instruction might need 300–800 tokens; a detailed procedure might need 1,000–3,000. These are starting examples, not fixed skill requirements. Three loaded instructions of 700 tokens use 2,100 before documents and answers.

You might initially reserve 300–800 tokens for a short answer or 1,500–4,000 for a detailed explanation or code. Actual needs depend on the task; reasoning and tool output may require more. Page counts are unreliable: language, tables, code and the tokenizer affect the count.

These ranges help you begin; they are not a reason to fill every field with the largest number. A larger budget alone does not improve answers. Replace estimates with counts of your own material as you learn.

Examples of a complete budget

The table shows four practice tasks, with every allocation, its total, the chosen window and spare space. A simple chat may need no tools or skills. Research gives more space to documents; an agent needs room for later actions.

These token allocations are examples, not hardware recommendations. Pick a similar task, substitute your numbers and check memory. Workload buttons may produce different values because balancing considers your settings and locks.

Count your material and estimate next steps

On the context page, choose Count text. Paste text or select a supported text file, choose its budget category and review the result before applying it. For PDF, extract the text first: file size in megabytes is not a token count.

Read the result label. Without the matching tokenizer, the count is approximate. If an exact-count service is configured, text may be sent to it; the local version uses the website’s local service. Counting an isolated text does not automatically include the full request’s formatting.

The step simulator estimates continuation: four steps of 500 tokens means 2,000. It takes the larger of that and the existing future-step reserve, avoiding double counting. It neither runs the agent nor predicts how many actions it will take.

Two meanings of “hybrid”

Hybrid architecture describes the model’s internal design. Different layers may store information differently. Such a model need not run on both CPU and GPU, and “hybrid” does not mean broken.

Splitting work between GPU and RAM describes where the model is placed. This can also apply to other architectures. RAM provides room, but processing through the CPU may be slower. We estimate capacity, not speed.

Catalog models have architecture profiles. The result remains a text-work estimate; image and audio processing are not included. Start with ordinary model selection and budgeting. Open How this estimate works when you want the calculation details.

Extra settings: when to revisit them

The runtime, such as Ollama or llama.cpp, actually loads and runs the model. Green Zone prepares a plan for it. Select yours if you know it; the website installs no software and checks no drivers.

The Key-Value Cache (KV cache) holds intermediate information about processed text. Working buffers support calculations. Other applications and the safety reserve leave memory outside the model’s work. They explain why the entire capacity cannot simply be assigned to the weight file.

Parallel requests mean several conversations processed simultaneously. Start with one. Sequential agent actions are not simultaneous requests. Change compact window cache, state copies and Flash Attention only when you understand your runtime settings; refining the estimate requires information from a real run.

Keep a configuration for later

Settings carry between tools and are saved in this browser. Pinning helps compare configurations; saving and exporting keep a configuration or report. Import restores values from a saved file.

The report contains a plan and assumptions, not computer measurements. Save important configurations before clearing browser data. To run the model, transfer the settings into your runtime and try a small task.

If the result does not meet your needs

The file is smaller than memory but the whole calculation does not fit? Try less context or a smaller file; the model needs working room too. RAM is available but unused? Check the RAM permission switch, capacity and system reserve.

An empty catalog? Remove filters and inspect the carried-over budget. Balance cannot help? Inspect locked allocations and the context lock. Text fits but memory does not? Return to the first calculator: these are separate checks.

Change one condition at a time and compare. It is safe to explore here: the site does not change model files or settings on your computer.