[1] Best as I can tell, the “local” version is the same as “CLI”, and “CLI” simply means using a cloud model, which you have to pay for. It also looks like JetBrains is hosting some models in their cloud which are free to use until December.
wannabe44 13 hours ago [-]
Slopped page. Apparently the target audience doesn't mind. Every sentence feels like a bland attempt at a punch lines rather than communicating something. Expected better from jetbeans.
gapo 1 days ago [-]
So this is what they were burning their cash on. Quite a few people did suggest this.
Interesting to see the min requirements. If they manage to bring that down further - this could be a huge turning point .
variety8675 1 days ago [-]
Watching the downfall of Jetbrains has been very sad
ternaryoperator 1 days ago [-]
What downfall? They continue to make great tools and I gladly pay for them.
simon_void 1 days ago [-]
why is Junie local only available for Mac & Windows? Why not Linux?
karmakaze 1 days ago [-]
And what's with needing a custom runner? Why not publish the model on Hugging Face so we can use it on any compatible with the model format?
browningstreet 1 days ago [-]
Look again. Read more carefully.
simon_void 1 days ago [-]
I've read it again: Junie CLI has a Linux version, Junie Local doesn't.
jt2190 1 days ago [-]
The GitHub page for the Junie [1] script explicitly says "MacOS/Linux": It's the same curl command for the CLI
There is a note in the bash script [2] that the local model needs MacOS for some reason:
install_local_model() {
echo ""
log "Junie is installed. Setting up the local model..."
if ! curl -fsSL "$LOCAL_MODEL_URL" | bash; then
# The local model is an optional extra on top of a finished Junie install,
# and it is the part most likely to be refused -- it needs macOS 26 on an
# Apple M5 or newer with 40 GB of RAM. Whatever went wrong here, Junie itself
# is installed, so say so plainly instead of leaving the run looking failed.
echo ""
log_error "The local model setup did not complete."
log "Junie itself is installed and ready to use -- run: junie"
log "To try the local model again: curl -fsSL $LOCAL_MODEL_URL | bash"
exit 1
fi
}
But the GitHub repo README for "local" [3] does have instructions for Linux though, and makes it seem like the --local flag will work. As a hardy Linux hacker, you should probably just try it and see what happens.
Edit: Interestingly, I got Junie CLI on my M4 Mac but did NOT get the local model installed, which t looks like iwas going to me Qwen3.6-27B-MLX-4bit. (Needs a M5 GPU apparently.)
I did start Junie and it prompted me to connect to a local model, successfully detected Ollama and connected to that.
Anyone on HN have real experience with their agent harness?
I've mostly used codex in the past (and in IDE plugins) with llama.cpp self-hosted models. Aside from packaging those things directly together (or another inference engine) does this bring something novel? Aside from a marketing angle.
> Fully local: Nothing leaves the machine
> Offline
> Free to use: No card, no subscription, no credits. [1]
https://junie.jetbrains.com/local
I’ll definitely give this a try.
[1] Best as I can tell, the “local” version is the same as “CLI”, and “CLI” simply means using a cloud model, which you have to pay for. It also looks like JetBrains is hosting some models in their cloud which are free to use until December.
Interesting to see the min requirements. If they manage to bring that down further - this could be a huge turning point .
Edit: Interestingly, I got Junie CLI on my M4 Mac but did NOT get the local model installed, which t looks like iwas going to me Qwen3.6-27B-MLX-4bit. (Needs a M5 GPU apparently.)
I did start Junie and it prompted me to connect to a local model, successfully detected Ollama and connected to that.
[1] https://github.com/JetBrains/junie [2] https://raw.githubusercontent.com/JetBrains/junie/refs/heads... [3] https://github.com/JetBrains/junie/tree/main/local
Anyone on HN have real experience with their agent harness?
I've mostly used codex in the past (and in IDE plugins) with llama.cpp self-hosted models. Aside from packaging those things directly together (or another inference engine) does this bring something novel? Aside from a marketing angle.