Head of Hardware at JetBrains. Any sufficiently advanced technology is indistinguishable from magic.
Munich, Bavaria
- Tweets 4,092
- Following 118
- Followers 28,212
- Likes 5,108
A colleague at JetBrains built a cool model based on Qwen 3.6 and 3.8 27B
Essentially, he combined the two models and ended up with a model that performs better than 3.6 and is significantly faster than 3.8 when run locally.
I ran this model all last week, and it performs very well in real-world scenarios.
The model is available for free on HuggingFace
huggingface.co/collections/J…
3
10
90
7,156
Ivan Kuleshov retweeted
Today we introduce JetBrains Air - the product system for software development orgs in the age of agents.
Built for developers, team leads & engineering directors.
Open and flexible: steer and control almost any LLM & harness.
Read more:
jb.gg/air-announce
77
77
460
243,328
Ivan Kuleshov retweeted
Junie has a new home on X: @junie_ai 👋
Follow for new features and announcements.
First up: a smarter Junie Local, a new blended model, and experimental Windows support. The team has the details below 👇
We mixed Qwen 3.8 and 3.6. Literally.
A 50/50 weight merge. One 27B model. 71% fewer output tokens than 3.8 in our internal coding eval.
Junie Local has an update. Windows devs, you're invited too 🤝
4
6
95
24,565
Technical details blog.jetbrains.com/junie/202…
There's a lot of work behind this.
And the results are really great, especially for running it on hardware that isn't the fastest
5
1,453
There's a lot of work behind this.
And the results are really great, especially for running it on hardware that isn't the fastest
We mixed Qwen 3.8 and 3.6. Literally.
A 50/50 weight merge. One 27B model. 71% fewer output tokens than 3.8 in our internal coding eval.
Junie Local has an update. Windows devs, you're invited too 🤝
1
9
2,854
Ivan Kuleshov retweeted
JetBrains takes its first step toward Local AI.
MacBook M5 users can set up Junie Local in one click and get a solid agentic experience.
Unlimited & Fully local.
Folks also had some fun presenting it on a plane :)
Check it out & stay tuned for more:
junie.jetbrains.com/local
Junie can now run entirely on your machine, and it’s completely free to use.
One command: /local. We picked and optimized the whole stack around the model: agent harness, model parameters, and inference engine. No tokens, no quota, no code ever leaving your machine.
20
46
414
65,752
A project we've been working on lately. I mentioned it before under the name "Junie Cage" 😁
blog.jetbrains.com/junie/202…
1
14
1,515
Set up the entire environment to launch and use a local model with a single command directly from the agent.
The hardware options are currently very limited, but the list will expand in the coming weeks. We are testing various configurations right now
Junie can now run entirely on your machine, and it’s completely free to use.
One command: /local. We picked and optimized the whole stack around the model: agent harness, model parameters, and inference engine. No tokens, no quota, no code ever leaving your machine.
1
11
1,660
I think the biggest sign will be when they stop using model numbers. There will just be the model name, without a number. And the model will gain a certain kind of neuroplasticity and will keep improving day by day.
The day we develop really good models. There will be signs.
Reliability increasing despite load going up and up. Sudden efficiency gains. Things getting faster. Resets.
These kinds of things.
1
8
2,639
The new model from JetBrains. Works great on GPU
Mellum started with code completion.
Mellum2 is built for more – handling both natural language and code.
A 12B-parameter open-source LLM for routing, RAG, and sub-agents, optimized for ultra-low-latency inference.
Now on @huggingface.
Learn more: jb.gg/zpb9dp
1
1
25
6,013
The perfect use case for
nvidia.com/en-us/products/rt…
Introducing the Lamark agent,
The agent that actually trains the model for you.
github.com/merocle/lamark-ag…
It's designed with a focus on the “Local First” philosophy.
I tested and developed it primarily for the Nvidia DGX Spark to leverage its strengths - specifically, its ability to fine-tune models. The agent learns every night through conversations with you.
The focus on tools and the ability to call cloud AI models is implemented as a tool, too.
The agent is completely open-source
The next step will be to add support for more devices, and of course, the new @NVIDIAAI RTX Spark.
It is based on Hermes and supports all of its features. Although personally, I've mostly used Telegram and the CLI
I'm Inviting Contributors!
1
4
2,144
Introducing the Lamark agent,
The agent that actually trains the model for you.
github.com/merocle/lamark-ag…
It's designed with a focus on the “Local First” philosophy.
I tested and developed it primarily for the Nvidia DGX Spark to leverage its strengths - specifically, its ability to fine-tune models. The agent learns every night through conversations with you.
The focus on tools and the ability to call cloud AI models is implemented as a tool, too.
The agent is completely open-source
The next step will be to add support for more devices, and of course, the new @NVIDIAAI RTX Spark.
It is based on Hermes and supports all of its features. Although personally, I've mostly used Telegram and the CLI
I'm Inviting Contributors!
2
2
33
4,790
It’s finally up and running!
And the result is pretty decent.
Now I want to put this 1U platform... 😅
19
8
191
86,695
Some real-world test results.
1. 2-node model – simple coding
2. Single-node model – similar task
3. Single-node model – simple chat (Hi, what can you do?)
4. 2-node model – simple chat
As for the software @exolabs – it delivered the best results of everything I tested (but the choice for Mac OS isn’t great)
A 4-node cluster is next.
What’s really cool is that loading the model is significantly faster than on Spark (takes 15–20 seconds, rather than 5–10 minutes). But even with simple chats, the laptop starts making an obscene amount of noise.
8
3
128
12,043