| --- |
| license: apache-2.0 |
| tags: |
| - llama.cpp |
| - gguf |
| - local-ai |
| - inference |
| - mixture-of-experts |
| - vulkan |
| - rocm |
| - metal |
| - text-to-image |
| - text-to-speech |
| - text-to-audio |
| - text-to-video |
| --- |
| |
| # Kera |
|
|
| **A local AI engine and app, built on llama.cpp, that runs big models fast on your own computer.** |
|
|
| By Kera Labs. Version v0.1.1: it works and is measured on real hardware, but not every goal is reached |
| yet (below). |
|
|
| Kera runs the same GGUF model files as llama.cpp and Ollama, and gives exactly the model's own answers unless you |
| turn on a mode that trades a little quality for speed (each one's measured cost is published). What it adds: |
|
|
| - **It writes faster.** It guesses the next few words ahead by itself where that pays (copying from the chat, the |
| model's own MTP head), measures whether each guesser pays, and drops the ones that don't. |
| - **It picks the fastest back-end for each message.** On an AMD machine with the optional ROCm add-on, |
| `kera serve --backend auto` measures Vulkan and ROCm on each model and sends every message to the one that |
| finishes it sooner: chats to Vulkan, long documents (and their follow-ups) to ROCm. |
| - **Chats survive restarts.** Each chat's state is kept on the SSD, so a follow-up or a restart doesn't read the |
| whole conversation again. |
| - **Models bigger than memory.** A mixture-of-experts model that doesn't fit is streamed from the SSD, expert by |
| expert, with the experts it uses most kept in memory. |
| - **Several people at once.** `kera serve --multi-user` reads a new message before the rest of someone's long |
| document, keeps everyone writing while a long prompt is read, and reads a start several chats share only once. |
| - **It explains your computer.** The first time it starts, it measures your drive, memory and graphics chip in about |
| 30 seconds and says in plain words what you can run. |
| - **OpenAI- and Anthropic-compatible** server (`kera serve`), so existing chat apps and agents work with it. |
| - **Pictures, music, speech and video, on your computer too** (0.1.1). Vision models get their image reader by |
| themselves; `kera image` and `kera serve` run image models (SD, SDXL, SD3.5, FLUX, Qwen-Image, Z-Image) through |
| stable-diffusion.cpp, and `kera video` video models (Wan, LTX-Video); `kera music` makes songs with ACE-Step 1.5 |
| through acestep.cpp, lyrics included; `kera speak` turns text into speech with Qwen3-TTS. Each engine comes in the |
| download, and the models are yours to download. |
|
|
| ## Measured |
|
|
| On a Minisforum mini PC (AMD Ryzen AI Max, Radeon 8060S graphics, Windows), with Nemotron 3 Nano 30B-A3B (Ollama's |
| own 22.8 GB file, so every engine ran the same weights), October 2026: |
|
|
| | | Kera, Vulkan | Kera, ROCm add-on | llama.cpp | Ollama | |
| | --- | --- | --- | --- | --- | |
| | Writing speed, everyday prompts (tokens a second) | **66** | 54 | 56 | 41 | |
| | First word in a chat | 0.36 s | **0.32 s** | 0.38 s | 1.1 s | |
| | First word after a restart | **0.79 s** | 0.87 s | 3.0 s | 3.2 s | |
| | A 95,000-word document: first word | 296 s | **138 s** | 303 s | 144 s | |
| | Follow-up question on it | 0.43 s | **0.33 s** | 0.41 s | 2.2 s | |
|
|
|  |
|
|
|  |
|
|
| Every number has its source, settings and date in `docs/BENCHMARKS.md` in the source. |
|
|
| ### As accurate as the usual way |
|
|
| Speed doesn't cost accuracy. One model file (Qwen2.5-1.5B-Instruct) through Kera and through Ollama 0.40.15, |
| 200 questions with known answers (100 school maths problems from GSM8K, 100 science questions from |
| ARC-Challenge), greedy: **Kera 149 right, Ollama 147**; both give the model the same prompt, token for token. |
| Pictures made with SD-Turbo through Kera and through the model's own reference code (Hugging Face diffusers), |
| with the same seed and weights, look identical (46.7 dB alike). Details: `docs/QUALITY.md` in the source. |
|
|
| ## Downloads |
|
|
| | File | For | |
| | --- | --- | |
| | `downloads/kera-v0.1.1-windows-x64.zip` | Windows 10 and 11 (Vulkan: AMD, NVIDIA and Intel graphics, or the CPU) | |
| | `downloads/kera-v0.1.1-windows-x64-hip-addon.zip` | Optional: AMD's ROCm for Radeon 8060S (gfx1151), unzipped into the same folder | |
| | `downloads/kera-v0.1.1-linux-x64.tar.gz` | Linux (Vulkan or the CPU) | |
| | `downloads/kera-v0.1.1-macos-arm64.zip` | Mac with Apple silicon (Metal), macOS 13.3 or later | |
| | `downloads/Kera-vs-Ollama.dmg` | A Mac app that races Kera against Ollama on your Mac, live, with one button | |
| | `source/kera-v0.1.1-source.tar.gz` | The source code | |
|
|
| **Try it:** unzip, run `kera` (a page opens in your browser and explains your computer), then |
| `kera serve -m model.gguf` for chat apps (`http://127.0.0.1:8080`). `kera serve --help` lists Kera's options. |
|
|
| No model weights are included: Kera runs GGUF files you download (from Hugging Face, or the ones Ollama already |
| has), each under its own licence. |
|
|
| ## Status |
|
|
| This is a preview. Kera's goals each have a gate measured on real hardware; the Windows machine above has passed |
| some of its measurements, and the rest (the Mac, NVIDIA machines and the bigger models) are still to be measured. |
| The Mac version builds and passes its tests, but hasn't been measured on a Mac yet. |
|
|
| ## Licence |
|
|
| Kera's own code: Apache License 2.0, copyright Kera Labs (`LICENSE`). It is built on llama.cpp and ggml (MIT) and |
| bundles other open-source code under its own licences (`NOTICE`, and `THIRD_PARTY_LICENSES.txt` in each download). |
|
|
