Homicide Desk Local AI Model
ollama.exe and llama-server.exe on your PC — privacy, disk, hitch, and how to quit clean.
Homicide Desk uses a language model for one job: phrasing live suspect and witness replies from the sentence you typed and the evidence you presented. Cases, evidence, endings, and who can be guilty are handwritten. The model ships inside the game and runs on your Windows PC. That is why the depot is about 13 GB (SteamDB lists 12.01 GiB on disk, 10.82 GiB download). Nothing you type is supposed to leave the machine. Steam still lists broadband for install, updates, Cloud, and the overlay — not because the booth phones a studio server.
This page is the operator sheet for ollama.exe and llama-server.exe. Hardware numbers live on system requirements. Version history lives on patch notes. How to actually question someone lives on interrogation. Do not skip those for a magic setting that makes the model name a killer. It will not.
What the bundled server does
On Play, the launcher starts a local server beside the CRT window. Replies generate from handwritten case facts plus your free-text question. There is no dialogue tree. There is no cloud account for AI. Shu’la Lab’s store disclosure is local generation, no login, no outbound prompt. If a firewall prompt appears, it is the bundled process binding on localhost, not a reason to expose the booth to the internet.
First reply on a cold 4 GB machine can hitch. Recommended RAM is 8 GB. The desktop, City Atlas, and Forensics Lab are not the heavy part. Keep this wiki on a second monitor if you can spare the memory; alt-tab storms are how people think the model crashed when it is only loading. Intel HD 4000 is the GPU floor for the OS. The model cares more about CPU and RAM.
The model cannot see a chromatograph peak you never calibrated. It cannot open the sealed strongbox in Case #001. It cannot save you from Internal Affairs. Present a document. Watch the stress meter. If you push with nothing, they lawyer up — that is authored failure, not a bad token.
Quit leftovers (1.3.9) and black screens
Before 1.3.9, closing the window left ollama.exe or llama-server.exe running. Steam still showed an in-session game. RAM stayed allocated. A second launch could fail. 1.3.9 ties the server’s lifetime to the window. If you are already on 1.3.9 and Task Manager still shows the process after a clean quit:
- End Task on both executables once.
- Verify integrity of game files.
- Relaunch from Steam and quit again.
- Do not kill the process mid-save. Cloud is on.
- Report Windows version, GPU, and build number on Discord or to devs@shulalab.com.
Black screens on boot were a different defect: launch-day hotfix, then 1.3.4, then 1.3.7. A leftover AI process can look like a hang after Play. Confirm the depot includes 1.3.9, then follow the black-screen steps on the spec page if the CRT never appears.
A partial 13 GB download can fail silently. Verify files before you reinstall. There is no macOS or Linux build. Steam Deck compatibility is unknown. Proton is off the supported matrix. Family Sharing shares the license, not a second model copy you can stream to a friend. The game is single-player only.
Privacy, languages, and what not to fear
Typed questions stay local. Do not paste a case file into an online chatbot and call it research — that is how you leak a variant you have not finished. This wiki’s Case #001 page will not name a killer for the same reason.
UI language does not download a second model. Patch 1.3.8 localizes case text in all ten interface languages. English and Arabic have full audio. Changing Settings will not shrink the 13 GB folder. Details: settings.
Interrogation quality still depends on you. Two-line questions. Present the receipt. Offer a Break costs salary. Request Dossier costs salary. The model will happily deflect if you gave it nothing to contradict. Use the interrogation planner as a list, not as a script it can read.
If replies never arrive but the desktop works, watch Task Manager for ollama.exe. Missing process: restart on the current patch. Pegged CPU on first load: wait. Pegged forever with no text: verify files, then report. Do not delete the model folder by hand to “save disk” and expect the booth to work.
Buyer context for the $5.99 price and the 17 percent intro offer through 15 September 2026: review. Official URLs: links. First hour at the desk: Getting Started.
The model is a voice. The case is still yours.
Frequently Asked Questions
Quick answers for detectives still on the clock.
Does interrogation need the internet?
Replies are local. Steam still needs a network for install, updates, and Cloud. Nothing you type is supposed to leave the PC.
Why is the install about 13 GB?
The language model ships in the folder. The CRT desktop is not the bulk. SteamDB lists about 12 GiB on disk for depot 4935211.
Steam still says I am playing after I quit. Why?
On builds before 1.3.9 the bundled server kept running. Update, then check Task Manager for ollama.exe or llama-server.exe.
Can the model tell me who to charge?
No. Facts and endings are handwritten. Variants can change the guilty party. See The Overtime for the spoiler fence.
Is this the same as a cloud ChatGPT feature?
No. There is no AI login and no studio prompt server. The booth talks to a process on your machine.
First reply takes forever. Is it broken?
Often it is a cold load on 4 GB RAM. Wait through the first generation, then it should settle. 8 GB is the comfortable spec on system requirements.