Best AI tools to run on your own computer
Chat, coding, and voice tools that run on your machine. Your prompts stay there. A Tested badge means we ran that tool. Test pending means we only read the official page.
Short answer
Some grades are our best guess from reading about the tool. Once we test it ourselves, we swap in the real score. Published scores on this job: Gemma 4 (tested tier A, score 8); Kokoro (tested tier A, score 8); whisper.cpp (tested tier A, score 8); llama.cpp (tested tier B, score 7); Ollama (tested tier B, score 7); Piper (tested tier B, score 7); GPT4All (tested tier D, score 5). Research order: Flint (research tier B); Phonon (research tier B); Ornith-1.5-122B (research tier C). A tested score replaces that tool's research letter. A sponsorship does not change either one.
Run AI locally
Chat, coding, and voice tools that run on your machine. Your prompts stay there. A Tested badge means we ran that tool. Test pending means we only read the official page.
10 tools
- A
Passed all 8 checks on a computer with no graphics card, but slow because it thinks before answering. The bigger hosted models were not tested.
Audio checked by transcribing it back. Not judged by ear.
Accurate and fast with no dependency trouble. Synthetic audio only.
- B
Fast and correct on CPU. Compile step and shifting CLI flags add friction.
Runtime installs fast and runs well on CPU. Scored on the runtime, not the small models' poor reasoning.
Fast and intelligible on CPU. Naturalness not judged by ear.
Open weights that Ollama and the Recursive announcement say you can run locally. We have not run it.
Small speech model meant to run on a phone or a laptop. We have not run it.
- C
The model card says it serves on one 80GB GPU and includes an Ollama command. That is local, but it is not a normal laptop. We have not run it.
- D
Easy and fast, but the older catalog model only got 2 of 8 tasks right.
Filed tools
| Tool | Tier | Score | Time | Cost | Result |
|---|---|---|---|---|---|
| Gemma 4 Free | A | 8 | Test pending | Test pending | TestedPassed all 8 checks on a computer with no graphics card, but slow because it thinks before answering. The bigger hosted models were not tested. |
| Kokoro Free | A | 8 | Test pending | Test pending | TestedAudio checked by transcribing it back. Not judged by ear. |
| whisper.cpp Free | A | 8 | Test pending | Test pending | TestedAccurate and fast with no dependency trouble. Synthetic audio only. |
| llama.cpp Free | B | 7 | Test pending | Test pending | TestedFast and correct on CPU. Compile step and shifting CLI flags add friction. |
| Ollama Free | B | 7 | Test pending | Test pending | TestedRuntime installs fast and runs well on CPU. Scored on the runtime, not the small models' poor reasoning. |
| Piper Free | B | 7 | Test pending | Test pending | TestedFast and intelligible on CPU. Naturalness not judged by ear. |
| Flint Free | B | Test pending | Test pending | Test pending | Test pendingOpen weights that Ollama and the Recursive announcement say you can run locally. We have not run it. |
| Phonon Free | B | Test pending | Test pending | Test pending | Test pendingSmall speech model meant to run on a phone or a laptop. We have not run it. |
| Ornith-1.5-122B Free | C | Test pending | Test pending | Test pending | Test pendingThe model card says it serves on one 80GB GPU and includes an Ollama command. That is local, but it is not a normal laptop. We have not run it. |
| GPT4All Free | D | 5 | Test pending | Test pending | TestedEasy and fast, but the older catalog model only got 2 of 8 tasks right. |
Queued task. Ship a small feature in a fixture repo. Log time to a working diff, cost, and failures.
- OllamaFreeLocal open-weight model runtimeChat assistantsBTested
- PiperFreeLocal text-to-speechVoiceBTested
- KokoroFreeSmall open-weight text-to-speechVoiceATested
- llama.cppFreeC/C++ LLM inference for GGUFChat assistantsBTested
- whisper.cppFreeC/C++ speech recognitionVoiceATested
- Gemma 4FreeA chat model you can run on your own computerChat assistantsATested
- GPT4AllFreeAn app for running chat models on your computerChat assistantsDTested
- FlintFreeAn open coding model you can run on your own computerCoding agentsBTest pending
- Ornith-1.5-122BFreeAn open-weight coding model with an MIT licenseCoding agentsBTest pending
- PhononFreeTiny on-device text-to-speechVoiceBTest pending
FAQ
Best AI tools to run on your own computer
Some grades are our best guess from reading about the tool. Once we test it ourselves, we swap in the real score. Published scores on this job: Gemma 4 (tested tier A, score 8); Kokoro (tested tier A, score 8); whisper.cpp (tested tier A, score 8); llama.cpp (tested tier B, score 7); Ollama (tested tier B, score 7); Piper (tested tier B, score 7); GPT4All (tested tier D, score 5). Research order: Flint (research tier B); Phonon (research tier B); Ornith-1.5-122B (research tier C). A tested score replaces that tool's research letter. A sponsorship does not change either one.
What will the agent actually run?
Ship a small feature in a fixture repo. Log time to a working diff, cost, and failures. Published on this job: Gemma 4 (tested tier A, score 8); Kokoro (tested tier A, score 8); whisper.cpp (tested tier A, score 8); llama.cpp (tested tier B, score 7); Ollama (tested tier B, score 7); Piper (tested tier B, score 7); GPT4All (tested tier D, score 5).
Can a sponsor become the top pick?
No. Sponsored placements are labeled and sold separately. They do not set the tier, the score, or the research letter on this page.
Which tools here are free?
Verified free or freemium on this job: Gemma 4 (Free), Kokoro (Free), whisper.cpp (Free), llama.cpp (Free), Ollama (Free), Piper (Free), Flint (Free), Phonon (Free), Ornith-1.5-122B (Free), GPT4All (Free). Paid tools and tools with unverified pricing are left off that list.