Local-optimized 1B models are the fastest route to running advanced LLM behaviors on modest hardware. This GGUF release packages a MiniCPM5-derived “Thinking” checkpoint fine-tuned on Fable 5 V2 into quantized files ready for llama.cpp, LM Studio, Ollama, jan, KoboldCpp and other GGUF runtimes—so you get tool-calling and chain-of-thought style outputs locally without a heavy cloud footprint.
Key Capabilities
- Tool/function calling (V2 improved): stronger function-calling and tool-use behavior compared with the prior release, making it better for agent workflows and structured API interactions. This shifts the model from simple chat completions toward detachable tool-oriented responses.
- Thinking mode and long context: the GGUF embeds MiniCPM5’s chat template and supports up to 131,072 tokens in the upstream config.json; useful for long-form reasoning, RAG, and multi-document workflows (practical limits depend on runtime and hardware).
- Lightweight local deployment: recommended quant Q8_0 (~1.1 GB) and an F16 (~2.1 GB) full-precision file let you run the model on consumer-grade GPUs or CPU-first setups via llama.cpp and compatible runtimes.
- Practical sampling defaults: "Think" defaults (temperature=0.9, top_p=0.95) preserve chain-of-thought behavior; a "No Think" mode lowers temperature and disables thinking blocks for concise answers.
- Benchmarks vs base: small but measurable gains on safety/bench tasks—e.g., BFCL/API-Bank improvements and API-Bank jump reflecting better callable-API behavior.
Who it's for & Trade-offs
Great fit if you need a locally hosted, function-capable conversational model for coding assistance, tool-enabled agents, or long-context RAG on limited hardware. It’s also convenient for experimenting with chain-of-thought flows without cloud latency. Look elsewhere if you require frontier-scale performance, the highest factual accuracy on complex knowledge tasks, or strict production-grade safety filtering—1B models trade raw capability for efficiency. Also note usable context is constrained by your runtime and available RAM/VRAM despite the 128K config ceiling. The package is distributed under Apache-2.0 and inherits the MiniCPM5 provenance.
