pip install -U langgraph. Usable without the rest of LangChain.
Memory
The model dominates. 32B Q4 is 21.2 GB.
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: LangGraph is a low-level orchestration library for stateful, long-running agents. It is not a consumer chat app.
How we got it: langchain-ai.github.io/langgraph, read 2 October 2026. Install is pip install -U langgraph or uv add langgraph. It can be used without the rest of LangChain.
Checked:
Key fact: The graph’s own overhead is small next to the model. Size the machine with our VRAM formula, not with a LangGraph “system requirement.”
How we got it: 27B Q4 = 18.2 GB. 32B Q4 = 21.2 GB. Persistence and checkpoints are disk and RAM, not a second GPU.
Checked:
Key fact: OpenBot’s product page names LangGraph as an AG-UI endpoint you can register.
How we got it: copilotkit.ai/openbot, read the same day. That is how a graph becomes something a person messages.
Checked:
Key fact: A four-way LangGraph fan-out at about 8K tokens per branch needs about 29.8 GB on Qwen3-32B at Q4 and about 23.5 GB on Qwen3-30B-A3B at Q4 when all branches run at once.
How we got it: Formula result + 4 x f16 KV cache for 8,192 tokens (262,144 and 98,304 bytes per token, site KV cache guide). Parallel branches per the LangGraph Graph API docs; parallel requests per the Ollama FAQ.
Checked:
Methodology
explains the VRAM formula, the bytes per weight at each quantization, and the full source list.
LangGraph is LangChain’s library for durable agent graphs: state, human checkpoints, streaming, and a mix of handwritten steps and model steps. The overview is explicit that it is low-level orchestration, that you do not have to use the rest of LangChain, and that people who want a prebuilt loop should start higher up. It is built by LangChain Inc. There is a commercial LangSmith side for traces and deploy. None of that is a personal agent you download instead of ChatGPT dots. If you searched “build your own dots,” this is the developer version of that sentence, and it assumes you can write Python.
What actually uses memory
A graph node that calls a tool is a function call. A graph node that calls a model is the entire model. Checkpoints live on disk or in whatever saver you configure. They are not a hidden 20 GB tax. The mistake is buying hardware for “LangGraph” as if it were a product with a minimum GPU. Buy hardware for the model the nodes call.
For a local endpoint, run Ollama or vLLM and point the chat model in the graph at it. Put Open WebUI or OpenBot in front if a non-developer has to talk to the graph. OpenBot’s page lists LangGraph as one of the AG-UI endpoints it will register. CrewAI is the higher-level alternative when you want roles more than you want an explicit state machine. Use one or the other for a first project, not both.
Tiers for the model inside the graph
Graph you are building
Model
Q4 formula
Router plus a couple of tools
K2-Horizon-7B
6.2 GB. Fine while you learn the API. Brittle on messy tool schemas.
Multi-step tool agent you would trust with a first pass
Qwen3.8-27B or K2-32B
18.2 GB or 21.2 GB. 24 GB minimum.
Long state, many turns
Same model, more KV cache
36 GB unified, or check the calculator at your turn count.
Human-in-the-loop is a feature of the library, and you should use it. A local graph with shell tools and no interrupt is how a 27B model deletes a directory. That is an application bug. The GPU will not stop it.
Worked example: fan-out and growing state
Two graph patterns change the memory bill. The first is fan-out. LangGraph's Graph API docs say that when a node has several outgoing edges, "all of those destination nodes will be executed in parallel as a part of the next superstep", and Send objects returned from conditional edges do the same for map-reduce. Every branch that calls the model is its own request with its own KV cache.
Four branches of about 8,000 tokens each: formula result plus four f16 caches of 8,192 tokens
Model at Q4
One branch at a time
Four at once
Where four at once fit
Llama 3 8B
7.9 GB
11.1 GB
A 12 GB GPU (RTX 4070, RTX 5070, Arc B580), about 0.9 GB left
Qwen3-30B-A3B
21.1 GB
23.5 GB
A 24 GB GPU at the edge, about 0.5 GB left
Qwen3-32B
23.3 GB
29.8 GB
RTX 5090 32 GB or Mac Studio M5 Max 36 GB (about 32 GB usable), about 2.2 GB left
Against a default Ollama server the branches do not all hold memory at once. Ollama's FAQ says each model handles one request at a time unless OLLAMA_NUM_PARALLEL is raised, and that raising it multiplies the context memory by the number of parallel requests. Set it to the widest fan-out you actually run, and size the machine from the four-at-once column.
The second pattern is state that grows. A graph that appends every tool result to its message list sends the whole list to the next model node. At f16, each 1,000 tokens of history costs about 0.26 GB of cache on Qwen3-32B, 0.10 GB on Qwen3-30B-A3B and 0.13 GB on Llama 3 8B. Say a 20-step loop adds 1,500 tokens per step (an example figure, not a measurement). By the last step that is 30,000 tokens: about 29.1 GB in total on Qwen3-32B, which is past a 24 GB card, and about 23.2 GB on Qwen3-30B-A3B, which still fits one. Trim or summarize the history before each model node, or plan for the larger machine.
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. Per-token cache sizes come from the layer and KV-head counts on that guide.
A developer machine, not a kiosk
The Mac Studio M5 Max 36 GB (B0HGKSQMX6) is the comfortable single-user box for a 32B graph plus a browser of docs. A 24 GB RTX 4090 (B0BG94PS2F) is the discrete equivalent. Do not buy the EVO-X2 unless the graph is serving other people or a large MoE. LangGraph will not saturate 128 GB by itself.
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.