Developers fine-tune OpenBMB MiniCPM5-1B, release 657MB local thinking model

The model, named MiniCPM5-1B-Claude-Opus-Fable5-Thinking, is based on OpenBMB's MiniCPM5-1B (1.08B params, LlamaForCausalLM architecture, 24 layers, 128K context). Developers used conversation and reasoning traces from Claude Fable 5 for supervised fine-tuning, not weight-level distillation—since they lack access to Claude's weights or logits. The fine-tuned model retains MiniCPM5's native thinking template and tool call format, supporting Think and No Think modes. GGUF repo offers four quantizations: Q4_K_M (~657MB, smallest), Q5_K_M (~751MB), Q8_0 (~1.1GB, recommended default), F16 (~2.1GB). Works directly in llama.cpp, Ollama, LM Studio, Jan, KoboldCpp. One-command Ollama: ollama run hf.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF:Q4_K_M. Recommended Think mode sampling: temperature=0.9, top_p=0.95. For developers using local open-source models, this model offers a low-resource (as small as 657MB) thinking model option, but understand its capability boundaries: fine-tuning only transferred response format and style, not frontier reasoning or broad knowledge. No benchmarks or training datasets were released; capability claims are unverifiable. Additionally, training on Claude outputs may raise licensing issues; not clarified in the model card.