Introducing Lalakai Labs.
Lalakai Labs is a small research lab in California. We build language models that run on the computers people already own, along with the evaluation and tooling needed to use them well.
Most frontier AI runs in data centers and is reached through a subscription or an API. We think a capable model should also run on a standard laptop, keep your data on your machine, and show how it reasons.
Hunter 4B
Our models are built for reasoning, agentic coding, and everyday knowledge work, and they are sized to run on ordinary hardware. We also train models for defensive security work. Those models are trained on curated cybersecurity examples covering causal reasoning, attack patterns, and incident response, so a security professional can triage alerts or investigate incidents without sending logs to an external service.
Our models write out their reasoning before answering, so you can see how they reached a conclusion and where they went wrong.
Hunter 1.7B
We build for the hardware most people have: a MacBook used for coursework, a laptop used for writing and research, a machine with 8 to 12 GB of memory. Some of our users have more capable hardware, like a Mac Mini or an Nvidia DGX Spark, and we are building for them too.
Our users include CS students, writers, researchers, and security testers who want to probe their own applications and systems with a model that stays on their machine. We want our models to respond quickly to simple requests and to do real work when the task is hard.
How we trained them
We don’t train on scraped web data. We build curated datasets with explicit reasoning traces, security-focused examples, and verified code, and every example goes through several rounds of review.
Training uses QLoRA fine-tuning. During quantization, we use an importance matrix to preserve the weights that matter most for reasoning and security tasks, so smaller versions keep those abilities.
Try them
What's next: We will always be working on new iterations and improvements to our current models. We are also releasing larger model families, including models at 27B parameters and above. We are a small team, and we will release each model when it is ready.