A developer crammed Claude's 'style' into a 657MB model you can run offline

A community developer known as GnLOLot has released a 1 billion parameter language model that mimics the response style of Anthropic’s Claude Opus, and it runs entirely on local hardware with no cloud calls. The model, dubbed MiniCPM5-1B-Claude-Opus-Fable5-Thinking, is a fine-tuned version of OpenBMB’s existing MiniCPM5-1B base, and it’s available now in GGUF format for tools like Ollama and LM Studio.

The secret sauce isn’t true distillation. No one has access to Claude’s actual weights or logits. Instead, this is supervised fine-tuning on synthetically generated conversation traces from a Claude model, a process the developer describes as training on ‘Fable 5 data.’ The distinction is critical. The smaller model learns to imitate Claude’s formatting, tone, and stylistic tics, but a 1.08B parameter budget physically cannot absorb the deep reasoning capabilities of a frontier model. You’re getting the mask, not the mind.

The technical specs are solid and verifiable. It inherits a massive 128K token context window from its OpenBMB base and ships in four GGUF quantization levels. The smallest, Q4_K_M, clocks in at roughly 657MB, while the maintainer’s recommended default, Q8_0, is about 1.1GB. You can pull it directly with a single command: ollama run hf.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF:Q4_K_M.

The release comes with some notable gaps. There are no published benchmarks to support any capability claims, and the training dataset remains private, making the results unverifiable. The licensing is also a gray area. While the base model is Apache 2.0, the act of training on outputs generated by Claude raises questions the model card doesn’t bother to answer. For hobbyists who want a taste of Claude’s conversational style in an air-gapped environment, it’s a fun toy. Just don’t expect it to do your taxes.