Cicikuş v2-3B targets 4.5GB inference

This is the kind of niche local-model drop that gets attention because the hardware target is realistic, even if the marketing copy is doing far more work than the benchmarks.

• –The practical hook is simple: a 3B model that aims to stay usable on modest local hardware instead of demanding 24GB+ GPUs.
• –Under the hood, this looks more like a tuned Llama 3.2 3B stack than a fundamentally new architecture, with Unsloth, QLoRA, and SFT-style training doing the heavy lifting.
• –Prometech publishes baseline scores like MMLU 58.7%, BBH 48.5%, GSM8K 40%, and MBPP 50%, which gives developers something to inspect even if the “consciousness” framing is unverifiable.
• –The bilingual Turkish/English angle is more interesting than the sci-fi branding because that can matter for regional and edge deployments.
• –The biggest catch is licensing: the model is publicly downloadable, but the “license: other” and commercial-use restrictions make it harder to treat as a default open model for production work.