A community developer has published a 1B-parameter language model that runs entirely on local hardware, requires no API key, and makes no cloud calls. The model is a supervised fine-tune of OpenBMB's MiniCPM5-1B base, further trained on conversation traces generated by Claude Fable 5. It ships in GGUF format with four quantization tiers, the smallest weighing in at roughly 657MB.
How the Model Was Built
The base architecture is OpenBMB's MiniCPM5-1B, a dense 1.08B-parameter model with 24 layers, grouped-query attention, and a 131,072-token context window. OpenBMB itself reports state-of-the-art results within the 1B class on its own benchmarks. The base model already includes a native thinking template that toggles reasoning through a system prompt, offering both a "Think" and a "No Think" mode. The derivative model retains this template, along with MiniCPM5's tool-call format.
On top of that base, the developer applied supervised fine-tuning using data generated through conversations with Claude Fable 5. The model card describes this as being "further fine-tuned on Fable 5 data" to sharpen coding and instruction-following abilities. The GGUF card repeats the same phrasing: "post-trained on Fable 5 data."
Why This Is Not Classical Distillation
The method used here differs materially from classical model distillation. In true distillation, a teacher model's logits or internal weights are transferred to a smaller student. No one outside Anthropic has access to Claude's weights or logits. Instead, the developer generated text outputs and reasoning traces from Claude, then used those as training data for the smaller MiniCPM5 base. This is supervised fine-tuning on generated outputs, not weight-level distillation.
OpenBMB's own base model, by contrast, uses a documented distillation pipeline between its own teacher and student checkpoints. The practical implication for end users is straightforward: the 1B model learns to mimic Claude's response style and formatting. It does not inherit frontier-scale reasoning capability. A 1B parameter budget simply cannot hold the depth of a model orders of magnitude larger.
File Sizes and Runtime Compatibility
The GGUF repository offers four quantization options:
The frequently cited "657MB footprint" refers specifically to the Q4_K_M build, not the default. The model loads directly in llama.cpp, LM Studio, jan, and KoboldCpp without additional conversion. For Think mode, the repository recommends a temperature of 0.6, top-p of 0.95, and a repetition penalty of 1.05. The model may emit reasoning blocks before its final answer, which downstream applications can strip if needed.
Licensing and Open Questions
The base MiniCPM5-1B weights carry an Apache 2.0 license. However, training on outputs generated by Claude raises a licensing question that the model card does not resolve. Anthropic's terms of service place restrictions on using Claude outputs to develop competing models, and the legal status of fine-tuning an open-weights model on proprietary model outputs remains an unsettled area in AI governance. Users deploying this in commercial settings should treat that ambiguity as a live risk.
What This Means for the Local LLM Ecosystem
This release fits into a broader pattern. As frontier models grow larger and more expensive to run, the community is aggressively exploring how much capability can be squeezed into tiny parameter counts. The 1B class is becoming a proving ground for efficiency tricks: aggressive quantization, structured sparsity, and now teacher-generated fine-tuning data. The honest framing here matters. A 657MB model that can reason out loud and follow instructions is genuinely useful for edge devices, offline workflows, and privacy-sensitive applications. But buyers should not expect Claude-level performance. The value is in the format and the footprint, not in matching a frontier model shot for shot.
Watch for whether Anthropic or OpenBMB issue any statement on the licensing of Claude-derived training data, and whether this sparks a wave of similar community fine-tunes on other small base models.