A 1B model that thinks twice—when you ask it to

What it does

MiniCPM5-1B is a dense 1B-parameter Transformer released by OpenBMB for on-device and resource-constrained use. The repository hosts model weights, training recipes, and single-page cookbooks for deploying and fine-tuning across major inference backends. It also includes “Agent Skills”—reproducible workflow snippets—and a desktop pet demo that runs the model entirely locally.

The interesting bit

The same checkpoint serves both chat and reasoning modes via a built-in chat template toggled with enable_thinking, so you do not need separate models for quick answers and deliberate chain-of-thought. The project also publishes sparse-attention variants like MiniCPM-SALA for million-token contexts, though the current spotlight is on the 1B dense model.

Key highlights

• Claims 1B-class open-source SOTA with an average benchmark score of 42.57, versus a prior high of 35.61 in the same size class, with its best results in agentic tool use, code, and competition math.
• Ships with deployment and fine-tuning cookbooks paired with Agent Skills in ./skills/ to help developers reproduce workflows.
• Supports both “Think” and “No Think” chat modes from a single checkpoint.
• Available in multiple formats for local deployment: BF16, GGUF, and MLX.
• Includes a desktop pet demo driven entirely by the local 1B model.

Caveats

• Benchmark scores are project-reported averages across reasoning, knowledge, code, and other domains; independent verification is not discussed in the visible README.
• The repository is organized around published weights, deployment cookbooks, and agent skills rather than a from-scratch training framework.

Verdict

Developers building local assistants, edge agents, or low-latency tools should look here—especially if you want a single small checkpoint that can toggle between fast chat and slower reasoning. If you need large-scale cloud inference or training foundation models from scratch, this is not your stop.