Verity Mod Ollama Local AI Setup

What the Ollama option changes

Ollama runs a language model service on your computer. Stable Verity JE 5.7.4 on Forge 1.20.1 connects through a local LiteLLM gateway instead of using Groq for conversation. This removes the need for a Groq key for the language-model request, but adds hardware, storage, model, service, and configuration requirements. The 6.0.0 beta family changed this setup; its removal of LiteLLM does not apply to the stable instructions below.

Use Ollama only after the base Java mod works. Minecraft should launch with the correct loader, Verity should load in a clean profile, and the documented in-game interaction should appear. If the game crashes before the title screen, solve the loader or dependency problem first.

Release	Connection to prepare	Instructions to use
Stable 5.7.4	Verity to LiteLLM on 4000, then Ollama on 11434	Follow the gateway setup below
6.0.0 beta family	The maintainer removed the LiteLLM requirement	Use the 6.0.0+ tutorial and your installed beta’s settings
Legacy NeoForge 3.4.1	Separate older release	Do not copy the stable Forge fields from this page

Ollama’s native /api service and its OpenAI-compatible /v1 interface are different APIs on the same service. The stable gateway is a separate process. A port number that works in a browser is not enough to choose a Verity endpoint.

How the three processes fit together

On stable 5.7.4 a single typed message passes through three separate programs. Understanding the chain makes almost every failure easy to locate:

1. 1. Minecraft + VerityLiteLLM URL http://127.0.0.1:4000/v1/
2. chat request
3. 2. LiteLLMReads ollama/, forwards to Ollama
4. port 11434
5. 3. OllamaLoads qwen2.5:1.5b, writes the reply

1. Minecraft with Verity sends an OpenAI-style chat request to the LiteLLM URL, http://127.0.0.1:4000/v1/chat/completions.
2. LiteLLM receives the request, reads the ollama/ prefix to choose its Ollama provider, and forwards the request to http://127.0.0.1:11434.
3. Ollama loads qwen2.5:1.5b, generates the reply, and returns it back along the same path.

Each link can be tested on its own. If Ollama answers ollama run but LiteLLM’s --test fails, the gateway is the problem. If both answer but Minecraft is silent, the Verity config or the profile is the problem. Testing in this order prevents changing three settings at once and losing track of which change mattered.

Start the programs in the same order every session: Ollama first, LiteLLM second, Minecraft last. Closing the gateway terminal stops the middle link even though Ollama keeps running, which is one of the most common reasons a setup that worked yesterday appears broken today.

Check whether local inference fits your computer

Local models use disk space, memory, processor or graphics resources, and power. Larger models generally need more resources and may answer more slowly on modest hardware. The right model is one that Ollama can load reliably while Minecraft is also running.

Before installing a large model, review the current Ollama documentation and the model’s published size. Keep enough free disk space for the download and enough memory headroom for both applications. A system that can run Minecraft smoothly may still struggle when a model and a modded game compete for memory.

Start with a small model for the standalone service check below. This verifies Ollama, not that the model can reliably follow Verity’s gameplay instructions. Once the connection is stable, compare models one at a time.

Practical sizing guidance

Exact requirements depend on the model file, quantization, operating system, and graphics hardware, so treat the following as planning guidance rather than a guarantee:

Situation	Reasonable starting point	What to watch
Laptop or integrated graphics	The 1.5B default used by 5.7.4	Total memory use with Minecraft open
Desktop with a dedicated GPU	Try a larger model only after the default works	Graphics memory shared with the game
Older or low-memory machine	Consider Groq instead of local inference	Stutter, swapping, or very long replies

Minecraft’s own allocated memory still counts. If the launcher assigns a large heap to the game and the model also needs several gigabytes, the operating system may start swapping to disk. That shows up as frame drops in Minecraft and multi-second pauses before Verity replies. Reduce the game’s allocation to a sensible value, close browsers and recording software, and retest before concluding the model is too slow.

A larger model is not automatically a better companion. Smaller models respond faster, which matters for a horror experience where timing affects atmosphere. Change one variable at a time and note how long a typical reply takes in a real world, not just in a terminal.

Will a model fit alongside Minecraft?

The common failure with local AI is not that Ollama refuses to run; it is that the model and modded Minecraft together need more memory than the computer has, so the system starts swapping to disk. Minecraft stutters and Verity takes half a minute to answer. Estimate this before downloading a larger model.

Memory budget calculator. All values in GB. The starting values describe a 16 GB Windows PC running the 5.7.4 default model.

Total RAM in the computer Memory assigned to Minecraft in the launcher Model download size (from the Ollama tag page or ollama ls) Other open programs (browser, Discord, OBS) Operating system reserve

Estimated memory needed

—

Headroom left

—

Without JavaScript: needed ≈ system reserve + Minecraft allocation + 1 + model size × 1.25 + other programs. Keep at least 2 GB of headroom.

How the estimate works, so you can adjust it:

• Minecraft uses more than its allocation. The launcher setting caps the Java heap; the game process also needs memory for graphics and native code. The calculator adds about 1 GB for that.
• A loaded model is larger than its download. Ollama needs extra memory for the conversation context, so the calculator adds 25% to the download size. The exact figure depends on the model and the context length.
• Operating systems need a reserve. Windows commonly uses 3–4 GB at idle; macOS and Linux are often lower. Check Task Manager or Activity Monitor with nothing open if you want a precise number.
• A dedicated graphics card changes the picture. If the model fits in the card’s own memory, Ollama can run it there, which frees system RAM but competes with Minecraft’s rendering. On a laptop with shared graphics memory, use the calculator as written.

If the headroom is negative, do not raise Minecraft’s allocation to compensate; that makes it worse. Lower the allocation to what the modpack actually needs (a light Verity profile rarely needs more than 4 GB), choose a smaller model, or use Groq for this computer.

Install Ollama from its official source

Install Ollama for your operating system from the official download page. Avoid repackaged installers or bundles promising a preconfigured Verity model. Open Ollama, then open a terminal on macOS/Linux or PowerShell on Windows.

Use these documented Ollama CLI commands, one at a time:

ollama pull qwen2.5:1.5b
ollama ls
ollama run qwen2.5:1.5b "Reply with one short greeting."

Expected results: the download completes, ollama ls includes qwen2.5:1.5b, and the final command prints a generated greeting. The wording varies. This 1.5B model is listed in Ollama’s library and matches 5.7.4’s configured default model. That default is not a guarantee of hardware performance or gameplay quality.

If the CLI cannot connect to Ollama, start the installed app or run ollama serve in a separate terminal and leave it open. Do not start a second server if one is already listening. If ollama is not recognized, reopen the terminal after installation and check the platform’s installation instructions.

Record the exact model name displayed by Ollama, including any tag. Similar-looking names can identify different models or versions. “Model not found” often means Verity requests a name that is not installed, not that the mod itself failed.

Verify the local service

Ollama’s API documentation uses http://localhost:11434/api for its native local API. Check the service without Minecraft:

curl http://localhost:11434/api/tags

In Windows PowerShell, use curl.exe http://localhost:11434/api/tags to run the curl executable. The documented response is JSON with a models array. Look for the model you downloaded. An empty array means the service answered but has no listed models; connection refused means the service or address still needs attention.

This is a diagnostic URL, not a universal Verity config value. Ollama’s native /api routes, its OpenAI-compatible /v1/ routes, and a separate LiteLLM gateway are different entry points.

“Localhost” means the same computer as the process making the request. If Minecraft runs in a virtual machine, container, remote desktop environment, or separate server, localhost may point to the wrong machine. Keep the first test on one computer. Remote exposure introduces firewall and security decisions that are outside the simplest Verity setup.

Do not open the Ollama service to the public internet merely to fix a local connection. A local firewall rule should be as narrow as the setup requires. Follow official Ollama guidance for any non-local deployment.

Start the LiteLLM gateway for 5.7.4

Leave Ollama running. The LiteLLM quickstart currently recommends this isolated CLI installation with the uv package tool available:

uv tool install "litellm[proxy]"

The current package requires Python 3.10 or newer; the documented uv route can provision a compatible Python. If you already have LiteLLM, use litellm --version to check that the command is available. Installation alone does not start the gateway.

Start the gateway in a terminal and leave it open:

litellm --model ollama/qwen2.5:1.5b --api_base http://127.0.0.1:11434 --host 127.0.0.1 --port 4000

This combines LiteLLM’s documented Ollama model prefix with its CLI host, port, and API-base options. Expected result: a running proxy listening on 127.0.0.1:4000, with Ollama as the upstream service on port 11434. The explicit loopback host keeps this local test off public network interfaces.

In a second terminal, run the quickstart’s standalone request test:

litellm --test

Expect a generated chat-completion response. If it reports connection refused or a missing model, fix the gateway or Ollama first.

If port 4000 is already in use by another program, the gateway will fail to start. Either stop the other program or choose a free port and change Verity’s LiteLLM URL to match. Keep the two values identical; a gateway on 4001 and a config pointing at 4000 produces the same silence as a stopped gateway.

Configure stable Verity JE 5.7.4

The maintainer’s published 5.7.4 artifact defines these exact controls and defaults:

1. From Minecraft’s title screen, open Mods > Verity > Config. Expect Verity Configuration.
2. Open Local LLM and turn Use Ollama on. In this release it overrides AI Settings > AI Provider, even if that menu still says GROQ.
3. Set LiteLLM URL to http://127.0.0.1:4000/v1/ and Ollama AI Model to ollama/qwen2.5:1.5b. These are the shipped defaults and match the gateway command above.
4. Keep Thinking Mode (Qwen3 Models only) off for this Qwen2.5 test. Apply the changes and reopen Config to check them.

The stored field names are AISettings.use_ollama, AISettings.ollama_url, and AISettings.ollama_ai_model. Verity appends chat/completions to the configured base URL. Do not paste /api/tags into LiteLLM URL, replace port 4000 with 11434 while keeping these gateway settings, or remove ollama/ from this model field.

For a different installed model, change both the gateway’s --model and Verity’s Ollama AI Model, keeping the ollama/ prefix before the exact name from ollama ls. Leave speech services unchanged until typed replies work.

Keep the release boundary explicit: the older Ollama video is titled “up to 5.7.3.” The maintainer’s 6.0.0-beta.1 notes from July 29, 2026 announce a changed config menu and removal of the LiteLLM requirement. That beta change is not evidence that stable 5.7.4 can skip the older gateway setup. The currently linked 6.0.0+ tutorial is for the beta family, and beta.9 reworked AI configuration again.

Values that must match

Setting	Where it lives	Stable 5.7.4 value
Installed model name	ollama ls output	qwen2.5:1.5b
Gateway model	LiteLLM --model	ollama/qwen2.5:1.5b
Gateway upstream	LiteLLM --api_base	http://127.0.0.1:11434
Gateway listener	LiteLLM --host / --port	127.0.0.1 / 4000
Verity base URL	Local LLM > LiteLLM URL	http://127.0.0.1:4000/v1/
Verity model	Local LLM > Ollama AI Model	ollama/qwen2.5:1.5b

If any row disagrees with its neighbor, fix that row before testing again. Most “Ollama doesn’t work” reports trace back to one mismatched value in this table.

Stable 5.7.4 end-to-end test

1. Start Ollama and LiteLLM, and confirm both standalone tests answer.
2. Start the clean Verity JE profile.
3. Open a new test world.
4. Once Verity appears, open Minecraft chat with T, type Verity, say hello., and press Enter. Keep General > Immersive Mode off while testing.
5. Watch the Minecraft log and gateway output for the request.

A response in Ollama but not Minecraft suggests a Verity configuration or response-handling problem. No request reaching Ollama suggests the provider selection, endpoint, service availability, or firewall. A model-not-found response identifies the name or missing download. A very slow response may be resource pressure rather than a broken connection.

Wait for one request to finish. Sending many prompts can queue local work, consume more memory, and make the mod appear less responsive.

Reading the logs during a test

Keep three windows visible during the first real test: the LiteLLM terminal, the Minecraft launcher’s log or latest.log, and optionally a terminal for Ollama. Each one tells you how far a request travelled.

• Nothing appears in the LiteLLM terminal: Verity never sent the request to the gateway. Recheck Use Ollama, the LiteLLM URL, and that you launched the profile you configured.
• LiteLLM logs a request but reports an upstream error: the gateway reached for Ollama and failed. Check that Ollama is running and that the --api_base value is correct.
• LiteLLM logs a successful completion but Minecraft stays silent: the reply returned to the game. Look in latest.log for Verity errors around the same timestamp and confirm Immersive Mode is off for the test.
• The request takes a long time and then succeeds: the chain works. The delay is model loading or hardware load, not a configuration error.

When asking for help on the project’s community channels, share the relevant lines from these logs rather than a description. Remove usernames, local paths, and any keys first. See the crash and log guide for where Minecraft stores its log files.

Common Ollama failures

Service is not running: the model may be installed but no process is listening. Start Ollama using the current operating-system instructions and verify again outside Minecraft.

Exact model is missing: run ollama ls. For the stable 5.7.4 setup above, Ollama must list qwen2.5:1.5b; the gateway command and Verity field must both use ollama/qwen2.5:1.5b. Do not apply the gateway prefix to a beta’s direct Ollama connection.

Wrong endpoint: for this stable setup, Ollama listens on 11434 and the LiteLLM gateway on 4000. Verity’s base URL is http://127.0.0.1:4000/v1/. Watch for a stopped gateway, an accidental https scheme, or a remote hostname copied from another tutorial.

Profile mismatch: confirm that the Verity configuration belongs to the launcher instance you actually start.

Memory pressure: close heavy applications, use a smaller model, lower other game resource demands, and observe operating-system memory rather than repeatedly reinstalling.

Firewall or isolation: verify that the Java process can reach the local service. In managed environments, follow network policy rather than disabling security controls.

Gateway closed with its terminal: LiteLLM runs only while its terminal is open. Restart it with the same command before launching Minecraft each session.

For errors that mention authentication, rate limits, or Groq while Use Ollama is on, see API connection failed; those messages usually mean the provider override did not take effect.

Moving from stable to a 6.x beta

Keep the working stable profile unchanged while creating a separate beta profile. Verify that Ollama still answers the same standalone prompt, then configure the beta using its matching tutorial. Do not carry over ollama/ just because the stable gateway required that prefix: it selects LiteLLM’s provider, not part of Ollama’s installed model name.

What beta.9 actually needs, read from its jar:

1. In Mods > Verity > Config > AI Settings, set AI Provider to OLLAMA. The beta talks to Ollama directly at http://127.0.0.1:11434; there is no LiteLLM step and no port 4000.
2. Turn on Use Custom Model and type the exact tag Ollama shows in ollama ls, for example qwen2.5:1.5b, without the ollama/ prefix. The beta has no preset for Ollama; with the custom model empty it reports “Ollama requires a custom model name.”
3. Leave API Key empty; Ollama does not need one. Use AI URL under Advanced only if Ollama runs on another computer.

The beta also has a Thinking Mode switch. Leave it off on small local models: thinking makes replies slower and adds little at 1.5B.

After a successful typed reply, stop any unused old gateway only if no other application needs it. Keep Ollama running. If a later error still names port 4000, check whether you launched the stable profile or retained a gateway address. The stable test above is not a beta setup recipe. For multiplayer, remember that localhost belongs to the machine making the request, not automatically to the player hosting the world. See the multiplayer setup checks before moving a local service to another machine.

Text before speech

Keep the working Ollama settings when moving to microphone and voice testing. Ollama handles the conversation model; it does not establish that speech recognition or spoken audio works. In a compatible 6.x installation, JE Extra’s offline voice option changes speech output independently. A local chat model plus a cloud voice provider is still partly online.

Privacy and security boundaries

Local inference can reduce the number of prompts sent to a cloud model, but it is not a blanket guarantee that the entire game session is offline or private. Review the behavior of Minecraft, the launcher, Verity, other mods, and update tools. Avoid placing passwords, private documents, or sensitive personal information in prompts.

Keep the Ollama service local unless you intentionally design and secure a remote deployment. Do not expose it through a router, public tunnel, or unrestricted network interface as a quick fix. Logs can still contain prompts and local paths, so inspect them before sharing.

Decide between Ollama and Groq

Choose Ollama when your machine can run an appropriate model and you prefer local inference. Choose Groq when local resource use or model management is impractical and you accept a cloud provider. Neither option fixes a wrong loader, missing dependency, or unsupported Minecraft version.

Factor	Ollama (local)	Groq (cloud)
Account or key	None for inference	Free API key required
Hardware load	Shares memory and GPU with Minecraft	Minimal on your machine
Extra processes on 5.7.4	Ollama plus LiteLLM gateway	None
Prompt location	Stays on your computer	Sent to the provider
Works without internet	Model inference can	No

Keep a record of the working model name, Ollama version, service address, Minecraft version, loader, and Verity release. After any update, test Ollama alone, then Verity in a clean world, before returning to an important save.

Quick checklist

• Base Verity JE 5.7.4 loads on Forge 1.20.1 in a clean profile.
• ollama ls lists qwen2.5:1.5b and ollama run prints a reply.
• LiteLLM is running on 127.0.0.1:4000 and litellm --test answers.
• Use Ollama is on, with the LiteLLM URL and model matching the table above.
• A typed Verity, say hello. produces a reply in a new test world.
• Only then move on to voice, Immersive Mode, or a different model.