Crews and flows. Python >=3.10 and <3.14 per docs.
Shared server
One 27B Q4 is 18.2 GB, not 18.2 GB per role.
Enterprise
Separate hosted product. This page is the local library.
Checked
2026-10-13
Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.
Citeable facts
Key fact: CrewAI is a framework for agents, crews, and flows. Current docs require Python 3.10 or newer and older than 3.14.
How we got it: docs.crewai.com, read 2 October 2026. There is a separate Enterprise product. This page is the library on your hardware.
Checked:
Key fact: N agents do not require N copies of the model if they share one OpenAI-compatible server.
How we got it: Memory tracks resident weights. Role prompts are not extra models. 27B Q4 is 18.2 GB once, not 18.2 GB per role.
Checked:
Key fact: A 7B per role is usually worse than one shared 27–32B server.
How we got it: Quality judgment for local crews. Not a CrewAI benchmark. Formula: 32 × 0.5 × 1.2 + 2 = 21.2.
Checked:
Key fact: Three CrewAI roles sharing one Qwen3-32B at Q4 need about 27.6 GB with three 8K calls in flight, and about 23.3 GB with one.
How we got it: 21.2 GB formula + 3 (or 1) x 2.15 GB f16 KV cache for 8,192 tokens at 262,144 bytes per token (layer counts from the site KV cache guide). Ollama runs one request per model at a time unless OLLAMA_NUM_PARALLEL is raised (Ollama FAQ).
Checked:
Methodology
explains the VRAM formula, the bytes per weight at each quantization, and the full source list.
CrewAI is an open-source framework for agents, crews, and flows. The docs describe tools, memory, knowledge, and human-in-the-loop tasks, and they also sell an Enterprise console for deploy and triggers. This page is the library on a machine you own. Install notes on the docs site say Python must be at least 3.10 and below 3.14. Read the installation page when you build it. GitHub is the star-us link from those docs; use the docs as the source of truth if the README and the docs disagree.
One server, several roles
A crew is a cast list: researcher, writer, reviewer. Beginners hear “three agents” and buy three models. That triples VRAM and usually lowers quality, because each role is then a 7B. Point every role at the same local OpenAI-compatible endpoint — Ollama or vLLM — and give them different prompts. The weights load once. The KV cache grows with how many calls are in flight, which is a reason to keep the crew sequential unless you have memory to spare.
Open WebUI is the human oversight door. Let the crew write the first version. Read it before it sends anything. OpenBot is the alternative if you want those roles as chat channels instead of a Python process.
Model size per role, which is really model size once
Design
Resident Q4 weights
Verdict
Three roles, three 7B copies
About 3 × 6.2 GB if you actually load three
Spendy and still a 7B brain. Don’t.
Three roles, one 27B
18.2 GB
The default. 24 GB card or 36 GB unified.
One 32B, sequential crew
21.2 GB
Better writing and tool calls. Tight on 24 GB if two calls overlap.
A specialist 7B plus a 32B
4.2 + 19.2 + 2 = 25.4 GB in one server before cache (adding the two formula results, 27.4 GB, counts the 2 GB reserve twice)
Only if the small model does a tiny routing job. Otherwise drop it. Calculator is per model.
Flows in the CrewAI docs can pause for a person. Use that. A local crew that emails a customer with no checkpoint is a bug, not an architecture.
Worked example: three roles on one box
Take a crew of three roles (researcher, writer, reviewer) where each call carries about 8,000 tokens of prompt and context. Two costs add up: the weights you keep loaded, and a KV cache for every call running at the same moment.
Weights loaded, Q4, our VRAM formula
Design
Memory
Three different 8B models in one Ollama server
3 x 4.8 + 2 = 16.4 GB
The same three as three separate servers
3 x 6.8 = 20.4 GB
One Qwen3-30B-A3B shared by all three roles
20.3 GB
One Qwen3-32B shared by all three roles
21.2 GB
The +2 GB in the formula is the operating system and the inference engine, so it counts once per server. On that basis one shared 30B-class model costs about the same memory as three separate 8B servers, and every role gets the larger model.
Calls in flight at once: formula result plus one f16 cache of 8,192 tokens per call (1.07 GB for Llama 3 8B, 0.81 GB for Qwen3-30B-A3B, 2.15 GB for Qwen3-32B)
Model at Q4
One call
Three at once
Where it fits
Llama 3 8B
7.9 GB
10.0 GB
Three at once: a 12 GB GPU (RTX 4070, RTX 5070, Arc B580). One at a time: an 8 GB RTX 4060 at the edge.
Qwen3-30B-A3B
21.1 GB
22.7 GB
Both on a 24 GB GPU (RTX 3090, RTX 4090, RX 7900 XTX).
Qwen3-32B
23.3 GB
27.6 GB
One at a time: 24 GB at the edge. Three at once: RTX 5090 32 GB or Mac Studio M5 Max 36 GB (about 32 GB usable), about 4.4 GB left.
Totals add the cache on top of the formula result, as our KV cache guide does, so they lean high by up to one 8K cache.
Which column you are in is an Ollama setting. Ollama's FAQ says each model handles one request at a time by default (OLLAMA_NUM_PARALLEL=1), so a crew that fires three calls together gets them one after another and stays in the one-call column. Raise OLLAMA_NUM_PARALLEL to 3 and the calls run together, and Ollama multiplies the context memory by the number of parallel requests: that is the three-at-once column.
To point every role at the same server, CrewAI's LLM docs show LLM(model="ollama/<model>", base_url=...) with base_url set to the local Ollama server on its default port 11434, and the Ollama provider needs the LiteLLM extra (uv add 'crewai[litellm]'). The example model in those docs is llama3:70b. A 70B at Q4 is 44 GB by our formula, so it does not fit a 24 GB GPU or a 36 GB Mac; in the catalog it is at the edge of a Mac Studio M5 Max 48 GB and fits a 64 GB one (both Apple configure, no Amazon listing). Swap in a model from the table above before you run the example on a single card.
Hardware
A 24 GB GPU or the Mac Studio M5 Max 36 GB (B0HGKSQMX6) runs the shared 27–32B design. Step up to the EVO-X2 128 GB (B0F53MLYQ6) only when the crew also embeds a large corpus or you insist on a second full model. CrewAI Enterprise’s cloud triggers are a different bill and do not change this table.
This site uses cookies and shows personalised ads via Google AdSense. We and our partners store and access information on your device to serve relevant ads and improve your experience.
You can accept all cookies, decline (non-personalised ads only), or
manage preferences.
See our Privacy Policy.
We use Google Analytics
This site would like to use Google Analytics to understand how visitors use the site. If you accept, Google Analytics will load. If you decline, or if you make no choice, Google Analytics will not load. No ads are shown on this visit.
See our Privacy Policy and Cookie Policy.
Cookie preferences
Choose which cookies you allow. Strictly necessary cookies are always active.
Strictly necessary
Session state, security, and performance. Cannot be disabled.