JetBrains puts a price tag on local AI, and the lever is decomposition

JetBrains and DeepLearning.AI have released a free course, AI Coding Workflows: Hybrid to Local, built around PyCharm and its AI Chat ^[1]. What makes it worth an operator's attention is not the model list but the sequence: the course treats moving work off frontier APIs as a decomposition problem with metrics attached, rather than a stance about sovereignty ^[2][9].

The course starts on hosted Anthropic models in Claude Code, according to JetBrains, deliberately using familiar ground to introduce the techniques ^[3]. The premise is blunt: smaller models struggle with big, open-ended "vibe coding", and dividing and bounding the work is what keeps them on track ^[4]. JetBrains builds on its earlier spec-driven development course ^[5], and concedes that the example app is basic, arguing that this is the point, since a large model can do the upfront work of forming right-sized steps for a smaller one ^[6].

The mechanism is a subagent per roadmap phase, spawned fresh to manage context, with a cheaper model doing the implementing: Opus as the "big brain" for the main conversation, Haiku as the "little brain" for the work ^[7]. Every lesson closes with the change in tokens, turns, cost and estimated wall time ^[8]. JetBrains names four levers: specs shaped for the model size, specialist subagents to divide work, cheaper models for routine work, and metrics collected as evidence ^[9].

Stage two swaps the whole stack. The course moves to the OpenCode agent running inside PyCharm, adds OpenRouter as an inference router, and picks DeepSeek as the model ^[10]. JetBrains says the migration is straightforward because the UI is the same, and frames this as its position that IDEs should be open platforms for agents and models ^[11]. Note that OpenRouter is called out as a paid step ^[10]. The implementer becomes a dedicated Markdown subagent, with mandatory controls in the frontmatter and "persuasion" guidance in the body, running DeepSeek v4 Flash as the little brain ^[12]. JetBrains reports the metrics came out a lot cheaper than the Claude Code version ^[13].

Then the part that turns into a hardware requisition. Local inference is LM Studio serving Gemma 4 12B, targeted at a 32 GB laptop ^[14]. In the hybrid configuration, Gemma runs locally as implementer while DeepSeek v4 Flash is promoted to big brain, which JetBrains describes as working quite well ^[15]. Fully local puts Qwen 3.5 27B in the big-brain seat, and JetBrains reports results better than expected, crediting the guardrails ^[16].

The honest caveat is that this is a vendor course about the vendor's IDE, the results are self-reported, and the qualitative comparisons are published without the underlying token and cost figures ^[17]. The example app is admittedly trivial ^[6], so the interesting number for anyone budgeting is the one piece of concrete kit: a 32 GB machine ^[14].

What to watch: whether the metrics-first discipline holds on a real codebase rather than a teaching app, and whether the four levers survive when specs get messy ^[9]. JetBrains says more local AI material is coming and that it remains committed to human-in-the-loop workflows ^[18]. Also watch the model names, since JetBrains itself says the specifics change weekly and the ideas are the durable part ^[2].