Key fact: browser-use is the open-source library at github.com/browser-use/browser-use. The repository description is “Agents that use the browser.”
How we got it: GitHub page read 2 October 2026.
Checked:
By LLM Hardware team · Methodology ·
Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.
Key fact: browser-use is the open-source library at github.com/browser-use/browser-use. The repository description is “Agents that use the browser.”
How we got it: GitHub page read 2 October 2026.
Checked:
Key fact: Text-only browsing can use a 7–14B Q4 model (6.2–10.4 GB). Screenshot agents should be planned at 24 GB for a 27–30B multimodal model.
How we got it: VRAM formula. Qwen3.8-27B is 27B. Muse Glimmer is 30B and includes a perception encoder, per Meta’s August 2026 post.
Checked:
Key fact: The old OpenHands path docs.openhands.dev/sdk/guides/agent-browser-use was not re-confirmed after the docs split.
How we got it: OpenHands docs index now leads with Agent Canvas and the SDK. This page links the index, not an unverified deep path.
Checked:
Key fact: browser-use connects to a local Ollama model through ChatOllama. With no host set, the Ollama Python client uses OLLAMA_HOST or port 11434 on the same machine.
How we got it: browser-use supported-models page and browser_use/llm/ollama/chat.py (host defaults to None); ollama-python _client.py (_parse_host falls back to 127.0.0.1 and port 11434). Read 2026-10-30.
Checked:
Key fact: On a 12 GB RTX 5070, qwen3:8b at Q4 leaves room for about 34,400 tokens of f16 KV cache, and qwen3:14b for about 6,800.
How we got it: 8.2B and 14.8B (model cards) x 0.5 x 1.2 + 2 = 6.9 and 10.9 GB. Memory left divided by 147,456 and 163,840 bytes per token (2 x layers x KV heads x 128 x 2 bytes; 36 and 40 layers, 8 KV heads, config.json). Rounded down.
Checked:
Key fact: A 32,768-token browser run adds 5.4 GB of f16 KV cache on Qwen3-14B, 3.2 GB on Qwen3-VL-30B-A3B and 8.6 GB on Qwen3-VL-32B.
How we got it: 32,768 x 163,840, 98,304 and 262,144 bytes per token, from each config.json (40 layers and 8 KV heads; 48 and 4; 64 and 8; head_dim 128), site KV-cache method.
Checked:
Methodology explains the VRAM formula, the bytes per weight at each quantization, and the full source list.
browser-use is an open-source library that lets a language model drive a browser. It is also the kind of tool OpenHands and other agent scaffolds call when there is no API. There is no separate “browser-use appliance” to buy. You install the library, you run Chromium through Playwright, and you point it at a model. The commercial cousin people compare it with is Perplexity Comet, which brings a search index this library does not have.
A browser agent fails in two boring ways. The page changed, so the action is wrong. Or the model was swapped onto the CPU halfway through, so the loop takes minutes per click and you kill it. This guide can help with the second failure. It cannot publish a success percentage for the first. Sites change. A local model does not get a special exemption.
Decide up front whether the model reads extracted text or a screenshot. Text is cheaper and blind to layout. Screenshots need a multimodal model and a bigger card. Muse Glimmer’s research post is explicit that a perception encoder sits on top of the language-model weights. Do not budget Glimmer as “17 GB and done” if you are feeding it images. Meta’s envelope for language model plus KV plus perception plus the drafter is 24–32 GB.
Q4_K_M formula: params × 0.6 + 2. Chromium’s RAM is extra. On unified memory it comes out of the same pool. The calculator will not show the browser process.
| What the model sees | Weights | Card |
|---|---|---|
| DOM text, simple clicks | Qwen3-8B (qwen3:8b) 6.9 GB, Qwen3-14B (qwen3:14b) 10.9 GB, on the 8.2B and 14.8B in their model cards | 12 GB can hold the 14B weights. It is not roomy once Chromium is open on a shared-memory machine. |
| Screenshots | Qwen3-VL-30B-A3B (qwen3-vl:30b) 20.6 GB, Qwen3-VL-32B (qwen3-vl:32b) 21.8 GB, on the 31B and 33B Hugging Face counts | 24 GB discrete, or 32–36 GB unified. 16 GB is the wrong purchase. |
| Long task, many steps | Same weights plus f16 KV cache. 32,768 tokens add 5.4 GB on Qwen3-14B, 3.2 GB on Qwen3-VL-30B-A3B and 8.6 GB on Qwen3-VL-32B. | At 32,768 tokens Qwen3-VL-30B-A3B totals 23.8 GB, 0.2 GB under a 24 GB card. Qwen3-VL-32B totals 30.4 GB, past a 24 GB card and past a 36 GB Mac once macOS and a 4 GB browser pad are counted. On the 36 GB Mac Studio the 30B-A3B run leaves 4.2 GB. |
An editorial 1–4 GB on top of the formula, for the Playwright process and the agent scaffold, is a planning fudge. It is not an NVIDIA spec. The agent formula guide says the same thing in a table.
This table names one Ollama model per job and works out how much context fits beside it. Weights are the site formula at Q4 (parameters × 0.5 × 1.2 + 2 GB) on each model card's parameter count. Context is f16 KV cache from each model's config.json: 147,456 bytes per token for Qwen3-8B and Qwen3-VL-8B, 163,840 for Qwen3-14B, 98,304 for Qwen3-VL-30B-A3B and 262,144 for Qwen3-VL-32B. Token figures are the memory left after the weights divided by that per-token cost, rounded down. On a PC, Chromium runs in system RAM. On a Mac it shares the pool, so the Mac row also takes off 4 GB for macOS and the top of this page's 1 to 4 GB browser pad.
| Memory | Text agent (use_vision=False) | Screenshot agent | Verdict |
|---|---|---|---|
| 12 GB, such as the RTX 5070 | qwen3:8b at 6.9 GB leaves about 34,400 tokens. qwen3:14b at 10.9 GB leaves about 6,800. | qwen3-vl:8b at 7.4 GB leaves about 31,100 tokens. The 30B and 32B vision models do not fit. | A text-agent card. Use the 8B for long runs and the 14B for short ones. Screenshots only on an 8B vision model. |
| 16 GB | qwen3:14b leaves about 31,200 tokens, close to its 32,768-token native window. | qwen3-vl:8b leaves about 58,300 tokens. qwen3-vl:30b at 20.6 GB does not fit. | A 14B text agent with long runs. Still not a card for 30B-class screenshot models. |
| 24 GB, such as the RTX 3090 | qwen3:14b with its full 32,768-token window, 16.2 GB in all. | qwen3-vl:30b at 20.6 GB leaves about 34,500 tokens. qwen3-vl:32b at 21.8 GB leaves about 8,300. | The first tier for a 30B-class vision model. For long runs pick the 30B-A3B: its cache costs 37.5 percent of the 32B's per token. |
| Mac Studio M5 Max 36 GB (28 GB after macOS and a 4 GB browser pad) | Any of the text models above, with room to spare. | qwen3-vl:30b leaves about 75,200 tokens. qwen3-vl:32b leaves about 23,600. | Long screenshot runs on either vision model, with the browser in the same memory. |
The Qwen3-VL model cards describe a visual agent that "operates PC/mobile GUIs". That is the vendor's description of the job, not a success rate, and this page publishes no completion rate for any model. Ollama's files for these tags are smaller than the formula (qwen3-vl:30b is listed at 20GB), so the table leans high.
These steps follow the browser-use quickstart and its supported-models page, read on 30 October 2026. Install Ollama from ollama.com first. Then create a Python environment, install browser-use, and let it install Chromium for the agent:
pip install uv uv venv --python 3.12 source .venv/bin/activate uv pip install browser-use uvx browser-use install
On Windows the activate step is .venv\Scripts\activate. Start the Ollama server if the Ollama app is not already running, then pull a model in a second terminal. Pull the tag for your card from the table above: qwen3:8b or qwen3:14b for text, qwen3-vl:8b, qwen3-vl:30b or qwen3-vl:32b for screenshots. Ollama's library listed all five on 30 October 2026.
ollama serve ollama pull qwen3:8b
Save this as agent.py and run it with python agent.py inside the environment:
import asyncio
from browser_use import Agent, ChatOllama
async def main():
llm = ChatOllama(
model="qwen3:8b",
ollama_options={"num_ctx": 16384},
)
agent = Agent(
task="Find the number 1 post on Show HN",
llm=llm,
use_vision=False,
)
await agent.run()
if __name__ == "__main__":
asyncio.run(main()) Four settings carry the hardware decision:
For long tasks, browser-use's max_history_items setting keeps only the last steps in the model's memory; by default it keeps all of them. Capping it is the software side of the cache figures above.
If screenshot automation is the job, buy 24 GB, not 16 GB. The RTX 3090 holds qwen3-vl:30b with room for about 34,500 tokens of context. This listing is the Founders Edition of an Ampere-generation card, so check the seller and condition on the listing. The RTX 4090 this page linked before is listed as unavailable at our last catalog check, so it has no link here. On a Mac, the Mac Studio M5 Max 36 GB runs either vision model with the browser in the same memory. The RTX 5070 is the 12 GB text-agent card from the table: qwen3:8b with about 34,400 tokens, or qwen3:14b for short tasks.
RTX 3090 Founders Edition 24 GB on Amazon
ASIN B08HR6ZBYJ. Amazon Associates link. The live price is on that page, not locked in here.
Mac Studio M5 Max 36 GB on Amazon
ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.
ASUS TUF RTX 5070 12 GB on Amazon
ASIN B0DS6S98ZF. Amazon Associates link. The live price is on that page, not locked in here.
For a text-only agent, yes. qwen3:8b is about 6.9 GB at Q4 by our formula and leaves room for about 34,400 tokens of context on a 12 GB RTX 5070. Screenshot agents on 12 GB are limited to an 8B vision model such as qwen3-vl:8b; the 30B-class vision models need 24 GB.
browser-use stops sending screenshots and leaves out its screenshot tool, per its agent parameters page. A text-only model can then drive the browser from page text, without seeing the layout. The default is True.
Every step adds to the context the model keeps, and that context is held as KV cache. On Qwen3-VL-32B, 32,768 tokens of f16 cache are 8.6 GB on top of 21.8 GB of Q4 weights. browser-use's max_history_items setting caps how many past steps stay in that context.
Yes. Pass the other machine's address to ChatOllama with host, for example host="http://192.168.1.50:11434", and on that machine set OLLAMA_HOST to 0.0.0.0:11434 as the Ollama FAQ shows. The FAQ describes no login for the Ollama API, so keep that machine on a network you trust.
A correction or a hardware question goes to the contact page.