GnLOLot, a community developer, has published a 1B model that runs entirely on local hardware. model is MiniCPM5-1B-Cloud-Opus-Fable5-ThinkingBuilds for llama.cpp-compatible runtime with GGUF. It requires no API key and makes no cloud calls.
Proposed Model
The model is built on openbmb/MiniCPM5-1B. That base is an actual, documented release from OpenBMB. This is a condensed 1.08B-parameter model using a standard LlamaForCausalLM architecture. It has 24 layers, grouped-query attention, and 131,072-token context length. OpenBMB reports 1B-class open-source SOTA within its own comparison set.
Aadhaar already ships with a native thinking template. logic is toggled enable_thinkingOffering both think and no think mode. The derived model keeps that template and the tool-call format of MiniCPM5.
On top of that base, the developer implemented a fine-tune. The card states that the model has been ‘further refined on Fable 5 data to improve coding and adherence to instructions.’ The GGUF card rephrases this as ‘Post-trained on Fable 5 data’.
How is it actually made
The method described is not classical distillation. You don’t shrink the original model. Instead you generate multiple conversations with a teacher model. You capture fragments of its answers and reasoning as text. You then superimposed a small base model on those marks.
This distinction is important for accuracy. Classical distillation transfers the signal from the logits or weights of a teacher. No one has access to the cloud’s weights or logits. So this is supervised fine-tuning on the output generated, not weight-level distillation. OpenBMB’s own base model, in contrast, uses a documented on-policy distillation step between its own teacher and student checkpoints.
The practical effect is that the 1B model learns to mimic the response format and style. It does not absorb the inherent potential of the teacher. The 1B parameter budget cannot hold marginal-level logic.
Specifications that are checked
The context window is 128K tokens, inherited from base. config.json (131,072). The GGUF repository ships four quantizations. Q4_K_M is approximately 657MB and is labeled as having the smallest footprint. Q5_K_M is approximately 751MB. Q8_0 is approximately 1.1GB and is the maintainer’s recommended default. The F16 is about 2.1GB.
‘657MB footprint’ is the smallest amount, not the default build. The model loads directly into llama.cpp, Ollama, LM Studio, jan, and KoboldCpp.
Interactive: How the build works
The explainer below gives an honest breakdown of the build pipeline, footprint tradeoffs, and what a fine-tune can and can’t do.