Context window How much text a model can hold in view in one go.
Short answer. A context window is the maximum amount of text, counted in tokens, that a language model can take into account at one time: your instructions, the conversation so far, any documents you paste in, and the answer it is writing. Llama 3.1 models support up to 128K tokens, which is roughly 100,000 English words. A longer window does not make the model read better: it costs memory and time, and accuracy often drops for facts buried in the middle of a long input. Offices usually do better by retrieving the few relevant passages than by pasting whole files.
01
How it works
Everything the model uses to write the next token must fit in the window. That includes the system prompt, earlier turns, retrieved passages and the reply itself. If the total is too large, the software either refuses, or quietly drops the oldest text, and the model then behaves as if it never saw it. A model has a maximum window set by its training, and the software running it picks the size actually used.
That second part catches people out. Ollama’s documentation lists default windows of 4k tokens on machines with under 24 GiB of video memory, 32k with 24 to 48 GiB, and 256k with 48 GiB or more (as of October 2026). A 12-page contract of about 6,000 words is roughly 7,800 tokens and would not fit in a 4k window, however large a window the model supports.
02
Window sizes of common open models
Model
Context window
Roughly in English words
Llama 3.1 (8B, 70B, 405B)
up to 128K tokens
about 100,000 words
Qwen3-32B
32,768 tokens, or 131,072 with YaRN
about 25,000, or 100,000 words
Gemma 3 (4B, 12B, 27B)
128K tokens
about 100,000 words
gpt-oss-20b and gpt-oss-120b
131,072 tokens
about 100,000 words
Gemma 3 1B
32K tokens
about 25,000 words
03
What a long window costs
Memory: the KV cache grows with every token. For Qwen3-32B it is about 262 KB per token, so a full 131,072-token window needs about 34 GB for one conversation.
Time: the model must read the whole prompt before the first word of the answer appears, so a long prompt means a long wait.
Accuracy: in the “Lost in the Middle” study, models answered best when the relevant passage sat at the start or end of the input and worse when it was in the middle, even for models built for long inputs.
Noise: every irrelevant page you add gives the model more chances to follow the wrong passage.
04
Practical rules for an office
Count tokens, not pages. About 1.3 tokens per English word is a fair guide; Hindi and other Indian scripts can need several times more.
Leave room for the answer. A window that is 95% full of prompt leaves little space for the reply.
Prefer retrieval over stuffing. Search the documents, then give the model only the best few passages with their sources.
Check what the runtime actually allocated. In Ollama the ollama ps command shows the context length of a loaded model.
Long meetings: an hour of talk at 150 words a minute is about 9,000 words, or about 12,000 tokens, which fits easily in a 32K window.
05
Where ANVE stands
ANVE’s document question answering works over the office’s own documents and shows the source passage beside each answer, so a person can check it without rereading the whole file. Each meeting action item must quote the transcript line it came from.
06
Questions people ask
How many pages fit in a 128K context window?
About 100,000 English words, or around 200 pages at 500 words a page, before you subtract room for instructions and the answer. Scanned or table-heavy pages use more tokens per page.
Does a bigger context window mean better answers?
No. It means more text can be supplied. Studies show accuracy can fall when the key fact is in the middle of a long input, and the memory and time costs rise.
What happens when the context window is full?
It depends on the software. Some return an error and others drop the oldest text without telling you, so test your tool with a long input.
Is the context window the same as the model’s memory?
No. The model keeps nothing between requests. What looks like memory is earlier text that the software sends again with each new request.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
