How the computer setups work
Every Ollama + Hermes guide on this page is the same four steps:
- Install Ollama from ollama.com/download.
- Set the context to 64k. Open Ollama's Settings and move the Context length slider to 64k. Hermes needs the room, and Ollama otherwise starts at 4,096 tokens on GPUs under 24 GB. (On Linux the guide gives one command instead.)
- Download the model with the one
ollama pullcommand in the guide. - Connect Hermes. On Windows and Mac, use the free Hermes Desktop app. Ollama installs it if needed and connects it to the model:
ollama launch hermes-desktop --model <the model from the guide>To switch models later, run the same command with the other model's name. Prefer the terminal (or on Linux)? Use ollama launch hermes --model <the model from the guide> instead.
Each guide ends with a short task you paste into Hermes, which has it create and review a file, so you can see the agent really works.
Why some tiers use llama.cpp instead. Ollama cannot download a model from Hugging Face when the file is split into several parts. The biggest builds on this page are split, so their guides use llama.cpp's server, and Hermes connects to it in the same way. DeepSeek is too big for any single computer, so it runs on a rented GPU server.
| Tier | Model | Download | Runs with |
|---|---|---|---|
| 1 · Smartphone | LFM3-1.7B Uncensored | 1.13 GB | PocketPal AI |
| 2 · Entry-level laptop, 8 GB (GPU or unified memory) | Qwen3.5 4B Amadeus | 2.8 GB | Ollama + Hermes |
| 3 · Gaming PC, 12–16 GB GPU, or 16 GB Mac | Gemma 4 12B Uncensored | 6.9 GB | Ollama + Hermes |
| 4 · 24 GB GPU or 32 GB Mac | Qwen3.8 27B Uncensored v3 | 16.6 GB | Ollama + Hermes |
| 5 · 48–64 GB machine | Nemotron 3.5 Bolt 74B Uncensored | 29.1 / 51.7 GB | Ollama or llama.cpp + Hermes |
| 6 · 128–256 GB workstation | Qwen3.8 Flash Next Uncensored · GLM-5.3 Flash Abliterated | 98.5 GB · 109–200 GB | llama.cpp + Hermes |
| 7 · GPU cluster / cloud | DeepSeek V4.1 Flash 1017 Uncensored | ~355 GB, in 48 parts | SGLang on 4 × H200 + Hermes |
Tier 1 · Smartphone: LFM3-1.7B Uncensored
Hardware: an iPhone (iOS 15.1 or newer) or an Android phone. Download: 1.13 GB.
AtomicChat's uncensored build of Liquid AI's LFM3-1.7B is small enough to run fully offline on a phone. Install the free PocketPal AI app, add the model from Hugging Face inside the app, load it and chat. Nothing leaves your phone once the file is downloaded.
- Phone setup guide: private chat with LFM3-1.7B Uncensored in PocketPal
- Model page: LFM3-1.7B Uncensored
- Want the same model as an agent on a computer? Hermes + Ollama guide for LFM3-1.7B Uncensored. This one is an experiment, because a 1.7B model struggles with multi-step agent work.
Tier 2 · Entry-level laptop with 8 GB (GPU or unified memory): Qwen3.5 4B Amadeus
Hardware: a laptop with an 8 GB GPU, or a Mac with 8 GB of unified memory. Download: 2.8 GB.
Jitheswar's Amadeus is an abliterated build of Qwen's Qwen3.5-4B: the refusal behavior is removed without any new training. The publisher reports 1 refusal on its 304 test prompts, down from 231 for the original model, and slightly lower scores than the original on truthfulness and instruction-following tests. Those are the publisher's own numbers. It keeps Qwen3.5's thinking mode and image input.
Use the Q4_K_M file, the smallest in the repository. It leaves the most room on an 8 GB machine:
ollama pull hf.co/Jitheswar/Qwen3.5-4B-Amadeus:Q4_K_M
ollama launch hermes-desktop --model hf.co/Jitheswar/Qwen3.5-4B-Amadeus:Q4_K_MOn a Mac with 8 GB of unified memory the model shares that memory with macOS, so close other apps. There is also a 3.6 GB Q6_K file with higher quality, for machines with memory to spare. The fit on 8 GB is an estimate, not a YouRunAI measurement.
Follow the four steps in "How the computer setups work" above: install Ollama, set the context to 64k, pull the model, then launch Hermes. If the model runs out of memory with a 64k context, move Ollama's context slider down to 32k. The publisher's page does not report tool-calling tests for this build, and a 4B model can struggle with long multi-step agent tasks, so check what Hermes actually wrote.
Tier 3 · Gaming PC with a 12–16 GB GPU, or Mac with 16 GB unified memory: Gemma 4 12B Uncensored
Hardware: a gaming PC with a 12–16 GB GPU, or a Mac with 16 GB of unified memory. Download: 6.9 GB.
HauhauCS's Balanced uncensored build of Google's quantization-aware Gemma 4 12B. HauhauCS reports 0 refusals on its 465 test prompts. That is the publisher's own number.
ollama pull hf.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced:Q4_K_MTier 4 · 24 GB GPU or 32 GB Mac: Qwen3.8 27B Uncensored v3
Hardware: one 24 GB GPU (RTX 4090 or 3090) or a 32 GB Apple Silicon Mac. There is also a 10.2 GB build for 12 GB cards. Download: 16.6 GB.
InterpreticAI's uncensored Qwen3.8 27B, built on Nous Research's Arete-27B, with the remaining political censorship removed. This is the strongest model on this page that fits on one consumer GPU.
ollama pull hf.co/InterpreticAI/Qwen3.8-27B-Uncensored-v3-GGUF:Q4_K_MTier 5 · 48–64 GB machines: Nemotron 3.5 Bolt 74B Uncensored
Hardware: a 64 GB Mac or workstation for the 51.7 GB build, or a 48 GB machine for the 29.1 GB build. Download: 29.1 GB or 51.7 GB.
AtomicChat's uncensored build of NVIDIA's Nemotron 3.5 Bolt, a mixture-of-experts model built for local agents. It uses about 9B of its 74B parameters for each token, so it is fast for its size. AtomicChat reports 20 of 20 valid tool calls after its edit.
- 48 GB: the 29.1 GB AD-IQ2_XS build is a single file, so it runs in Ollama:
ollama pull hf.co/AtomicChat/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Uncensored-GGUF:IQ2_XS - 64 GB: the 51.7 GB AD-Q4_K_M build is the closest to the original model, but it comes as two files, so the guide runs it with llama.cpp.
- Hermes guide: Nemotron 3.5 Bolt 74B Uncensored (both builds)
- Model page: Nemotron 3.5 Bolt 74B-A9B Uncensored
Tier 6 · 128–256 GB workstations: Qwen3.8 Flash Next and GLM-5.3 Flash
Hardware: a Mac Studio or workstation with 128 GB of memory or more. Download: 98.5 GB (Qwen3.8 Flash Next) or 109–200 GB (GLM-5.3 Flash). Plan for a long download.
Two large mixture-of-experts models. Both guides use llama.cpp, because the GLM builds are split into several files and Ollama's support for Qwen3.8 Flash Next's new architecture with this community file is not documented.
- Qwen3.8 Flash Next Uncensored (177B, about 98.5 GB). OrcaRouter's abliterated build, with tools, thinking and vision. Fits a 128 GB machine with little to spare.
- GLM-5.3 Flash Abliterated (321B). Huihui's abliterated build of Z.ai's GLM-5.3 Flash: 109 GB for 128 GB machines, up to 200 GB for 256 GB machines.
Tier 7 · GPU cluster or cloud: DeepSeek V4.1 Flash 1017 Uncensored
Hardware: a rented server with 4 × NVIDIA H200 (564 GB of GPU memory), at least 256 GB of RAM and 700 GB of disk. Your own computer only runs Hermes and an SSH tunnel.
dealignai's uncensored DeepSeek V4.1 Flash runs with SGLang on the rented server. Hermes on your computer connects to it through a private SSH tunnel. Rent by the hour and stop the server when you are done.
If something goes wrong
- Hermes answers in chat but never reads or writes files. Make sure you started Hermes inside the folder the guide creates, and that
ollama psshows the model loaded with a 65,536-token context. Small models (tiers 1–2) fail multi-step agent work more often. Try the next tier up if your hardware allows it. - Ollama says the architecture is unknown. Update Ollama to 0.40.16 or newer and run the commands again.
- Out of memory, or very slow. Choose the smaller build in the guide, or lower the context to 32,768 tokens as the guide explains.
- Not sure what your machine can run? Save your hardware and YouRunAI rates every model for it.