What we're focused on.
Hunter 4B and Hunter 1.7B came out earlier this month, and since then the question we get most is what we’re doing next. Here’s the answer. There are no dates in it. Some of this is close, some of it is early, and one piece isn’t ready to show at all.
Fixing what’s weak in version one
Hunter 1.7B lists a 256K context window, but we trained it on sequences of 1,024 tokens and haven’t tested how it behaves beyond that. That gap is the first thing we want to close: train on longer sequences, then test the long-context claim properly instead of just repeating it.
The other known problems are on the model cards too. Hunter 4B can hallucinate and can be pushed into unsafe output with adversarial prompts. The 1.7B’s cybersecurity training set is only 2,500 examples, which is small. Both models get another pass on their data, and we’ll say what changed when they ship.
27B and up
We’re building larger model families at 27B parameters and above, for people with bigger machines. At 4-bit, a 27B model is roughly 16 GB of weights before you count context, so you want a Mac Mini with 24 GB or more, or something like an Nvidia DGX Spark.
Same rules as the small ones: it runs on your hardware, your prompts stay on your machine, and the weights are open under Apache-2.0. The difference is that a larger model should hold up on harder tasks, like long refactors or untangling a messy incident.
Security and agentic coding
Our models are built for two things: agentic coding in general, and defensive security work. On the coding side, we want models that can take a task, work through a codebase, use tools, and finish the job. On the security side, we want them to triage alerts, investigate an incident, and read through logs without anything leaving your machine. We plan to push both further, and the security side needs more curated examples and better incident-response scenarios than we have now.
One limit on the security side hasn’t changed. Those models are built for defense, and a qualified person should check what they say before anyone acts on it.
Benchmarks
Neither Hunter model has published benchmark results. Our evaluation has been internal. Every release after these will ship with benchmarks, plus what you need to rerun them on your own machine.
The coding agent
We want an agent that works in your project: it reads the codebase, edits files, runs commands, and checks the result, with a Hunter model doing the work on your machine. We’ve built part of it. It’s nowhere near ready, so it isn’t part of this release, and we’re not going to announce it until it is.
No dates
We set the bar for a release before we train, and a model that misses it doesn’t ship. A date on top of that only pushes us to lower the bar, so we don’t give one.