Small language models, built from scratch.
KaiLu is the KvcHub Labs family of chat models. Every checkpoint is trained from random initialization and sized to run offline on your own machine. KaiLu 3 Pro is the newest: KaiLu 3 grown to 1B parameters and trained further, open source under Apache 2.0, and runnable from a single 1.1 GB file.
- KaiLu 343.8 HellaSwag 36.7% · WinoGrande 50.9%
- KaiLu 3 Pronewest43.7 HellaSwag 35.4% · WinoGrande 52.0%
- KaiLu 2 2B43.3 HellaSwag 34.8% · WinoGrande 51.9%
- KaiLu 2 933M43.0 HellaSwag 35.0% · WinoGrande 51.0%
Average accuracy on HellaSwag (first 2,000 validation questions) and WinoGrande (all 1,267), measured by KvcHub with llama.cpp on each model's default download. Bars start at random guessing (37.5); gaps under about 1 point are within measurement noise.
Talk to KaiLu, right here
KaiLu runs inside your browser. The model downloads once and stays cached; after that, everything happens on your device and nothing you type is sent anywhere.
Load KaiLu 3 to start chatting.
Downloads once, then loads from your browser's cache. Best in Chrome, Edge, Firefox or Safari on a laptop or desktop.
- Speed
- –
- Last reply
- –
- Runs on
- –
- Model
- –
KaiLu is small: it makes mistakes, forgets earlier messages and works best in English, so check anything important. Powered by wllama (llama.cpp in WebAssembly); the browser builds are on Hugging Face.
The KaiLu family
Every KaiLu model is trained from random initialization. They borrow LFM2's hybrid layout of short-convolution and attention layers, not its weights.
-
Text Generation
KaiLu 3 Pro
KvcHub/KaiLu-3-Pro
KaiLu 3 grown to 1B parameters and trained further on web, math and instruction data, then tuned with supervised fine-tuning and DPO. One 1.1 GB GGUF runs it in Ollama, llama.cpp and LM Studio. Start here.
- Parameters
- 1.03B
- Layers
- 36 (24 conv, 12 attn)
- Context
- 1,024 tokens
- License
- Apache 2.0
-
Text Generation
KaiLu 3
KvcHub/KaiLu3
The 329M base model that KaiLu 3 Pro grew from. Pretrained on 6.6B tokens of web, textbook, math and code data, then tuned with supervised fine-tuning and DPO. The smallest and fastest KaiLu.
- Parameters
- 329M
- Layers
- 24 (16 conv, 8 attn)
- Context
- 1,024 tokens
- License
- Free for personal use
-
Text Generation
KaiLu 2 2B
KvcHub/KaiLu-2-2B-GGUF
The largest KaiLu, trained on a single Mac. It works through math, code and logic step by step before it answers.
- Parameters
- 1.90B
- Layers
- 104 (68 conv, 36 attn)
- Context
- 1,024 trained, 4,096 max
- License
- Apache 2.0
-
Text Generation
KaiLu 2 933M
KvcHub/KaiLu-2-933M
Built with MLX for Apple silicon, with a 546 MB GGUF build sized for phones.
- Parameters
- 933M
- Speed
- 260+ tok/s on Apple silicon
- Mobile RAM
- 1.15 GB
- License
- Apache 2.0
All KaiLu models work best in English and can state wrong facts confidently, so they aren't meant for high-stakes use. KaiLu 3 Pro and KaiLu 2 are open source under Apache 2.0. KaiLu 3 (329M) is free for personal use, study and modification; for business use, email [email protected] first.
Run KaiLu on your machine
Pick a model and a runtime. Commands come straight from each model card.
$ ollama run hf.co/KvcHub/KaiLu-3-Pro
Best in English, within 1,024 tokens of context, and at a low temperature such as 0.2, which the Ollama setup uses.
Choose a quantization
KvcHub/KaiLu-2-2B-GGUF- Download
- –
- Speed
- –
- Peak RAM
- –
- Perplexity
- –
Measured by KvcHub on an Apple M5 Pro: GPU generation speed, peak RAM at a 2K context, and perplexity on held-out text, where lower is better.
A quick look at Kava
Kava is the lab's pair of tiny token classifiers, at 636k and 2.01M parameters. They replace personal details with placeholders like [EMAIL] and [PHONE] before text is logged or sent to a language model. Try a browser version below; your text never leaves this page.
- Latency
- –
- Tokens
- –
- Memory (est.)
- –
- Sent over network
- 0 B
- [EMAIL]0
- [PHONE]0
- [CARD]0
- [IP]0
The browser demo uses lightweight pattern rules so it runs anywhere, with latency measured live on your device. The real Kava models do learned token classification in Python. Memory estimates assume fp16 weights plus the text buffers.