Someone Fine-Tuned OpenBMB’s MiniCPM5-1B on Claude Fable 5 Traces to Ship a 657MB Local Thinking Model

A developer going by GnLOLot just dropped a 1B-parameter model called MiniCPM5-1B-Claude-Opus-Fable5-Thinking, and the pitch is simple: it runs entirely on your machine, no API key, no cloud calls, GGUF files ready for llama.cpp and friends. The base is OpenBMB's MiniCPM5-1B, a documented 1.08B dense model with grouped-query attention and a genuinely huge 131,072-token context window. That part is real and verifiable.

What happened next is where the story gets interesting, and a little murky. GnLOLot took that base and fine-tuned it on what the model card calls Fable 5 data, supposedly improving coding and instruction-following. But this is not distillation in the technical sense people often assume. Nobody has access to Claude's weights or logits, so there's no way to transfer signal at that level. What actually happened is supervised fine-tuning on generated conversations — you prompt a teacher model, capture its text replies and reasoning traces, then train the small model to mimic that output style. OpenBMB's own base model, notably, does use a documented on-policy distillation stage between its own teacher and student checkpoints. This derivative skips that entirely and just imitates surface behavior.

That distinction matters more than it sounds. A 1B parameter model fine-tuned on someone else's outputs can pick up formatting habits, tone, maybe some conversational structure. It cannot absorb the underlying reasoning capability of a frontier model, because that capability isn't something text traces alone can transfer. The model keeps MiniCPM5's native thinking template, toggled via enable_thinking, so it can run in Think or No Think mode, and it recommends a temperature of 0.9 with top_p at 0.95 for the reasoning mode. None of that changes the ceiling on what a 1B budget can actually hold.

On hard specs, the numbers check out. The GGUF repo ships four quantizations: Q4_K_M at roughly 657MB (the smallest), Q5_K_M around 751MB, Q8_0 near 1.1GB and flagged as the maintainer's recommended default, and F16 at about 2.1GB. It loads in llama.cpp, Ollama, LM Studio, jan, and KoboldCpp, and there's a one-line Ollama command to pull the Q4_K_M build directly. Convenient, sure — but convenience isn't the same as capability, and there are no published benchmarks or training dataset to verify what this fine-tune actually improved.

There's also an unresolved licensing wrinkle worth flagging. The base weights are Apache-2.0, but training on outputs generated by Claude raises questions the model card doesn't address. Nobody's claiming this is illegal, but it's the kind of gray area that tends to matter once a project gets popular enough for someone to ask hard questions.