Information
Open weight fashions are having a second, pushed by management, selection, and value. Hybrid and native AI at the moment are getting severe seems to be, so JetBrains teamed up with DeepLearning.AI on a free AI Coding Workflows: Hybrid to Native course that covers the concepts and choices.
The course is now accessible and makes use of PyCharm and its AI Chat. Right here’s a peek into the course.
Claude Code: Subagents and cheaper fashions
We begin the course with, nicely, not-local. As an alternative, we use what you already know – Claude Code and its Anthropic fashions – to introduce among the strategies and “levers” that assist convey selection, management, and even value discount. (Sure, I wrote emdashes.)
We did a earlier course on Spec-Pushed Improvement (SDD) so in fact, we wished to start out there. Smaller fashions wrestle with large, open-ended “vibe coding.” Dividing and bounding the work retains smaller fashions on observe. Necessary observe: this course’s instance app is actually primary. You may say “that’s too straightforward.” However that’s a part of the takeaway: large mind fashions can do the upfront work, forming right-sized steps for smaller fashions.
We then illustrate this division with a Claude Code subagent. The primary chat immediate implements every roadmap section in a contemporary subagent, to raised handle context. This then provides the payoff: a less expensive mannequin for the implementer. Use a “large mind” (Opus) for primary dialog pondering and a “little mind” (Haiku) for implementation.
Every lesson finishes with metrics concerning the change in tokens, turns, value, and estimated wall time. Which brings us to the principle course purpose: studying the concepts as an alternative of the specifics, which change weekly.
New agent, inference, and mannequin
That covers the 4 levers:
- Specs formed for the mannequin dimension
- Specialist subagents to divide work
- Cheaper fashions for the routine work
- Gather metrics as proof to information pondering
The course then introduces selection and management:
- New agent: OpenCode
- New inference router: OpenRouter
- New mannequin and inference host: DeepSeek (through OpenRouter) by shifting to a brand new agent (OpenCode) utilizing inference routing (OpenRouter) to inference internet hosting and fashions (DeepSeek)
We first transfer to OpenCode, operating in PyCharm. JetBrains needs our IDEs to be open platforms for brokers and fashions. This makes the transfer from Claude Code to OpenCode easy: it’s the identical UI. We add OpenRouter (a paid step), join it to OpenCode, and select DeepSeek as a mannequin.
Subsequent we repeat our sequence: multi function chat, then context isolation utilizing a subagent. However this time, with a unique agent and mannequin.
We end by making a devoted implementer subagent in Markdown. This provides fairly plenty of levers of management: within the frontmatter for obligatory controls, and within the subagent physique for “persuasion” steering. Most significantly, we’ve the implementer use the smaller DeepSeek v4 Flash mannequin because the “little mind.”
In comparison with the Claude Code model, the metrics had been, unsurprisingly, quite a bit cheaper.
Hybrid and Native
Now for the principle attraction: for routine improvement, can we do some – and even all – of the work domestically?
We begin with a lesson on organising native AI: LM Studio because the inference server and Gemma 4 12B because the native mannequin, focusing on a 32 GB laptop computer.
We then configure the implementer subagent to make use of this native Gemma 4 mannequin, selling DeepSeek v4 Flash from final lesson’s “little mind” as much as “large mind.” The outcomes? Fairly good, because it seems.
Then the massive check: absolutely native, with Qwen 3.5 27B because the “large mind.” The outcomes: higher than anticipated, exhibiting that guardrails assist.
How did hybrid and native do? Each of those classes end with a overview of the metrics. That’s one of many large course takeaways: take a look at the proof. You possibly can see how small fashions wrestle, and see the impact of serving to them succeed.
Hybrid and Native AI Are Heating Up
A lot due to DeepLearning.AI each for working with us once more and for pushing to get this out quick. This matter is now red-hot within the information: Sovereign AI, privateness and safety, and naturally value. The improvements are coming actually quick and you will need to have a delicate introduction to the basics.
We’ll do extra updates right here on Native AI for management, selection, and value. Most of all, we at PyCharm consider within the human-in-the-loop. Keep tuned for extra on this.
