Research Vault

OpenBMB MiniCPM5-1B

Properties1
tagshuggingface, text-generation, coding, agentic, dense, instruct, on-device, tool-calling, long-context, local

3 min read

OpenBMB MiniCPM5-1B

One-line: 1B-class open-source SOTA on-device model (OpenBMB/Tsinghua NLP + ModelBest) — AA Intelligence Index 17.9, the top open-weight model at ≤1B params by 7.4 pts — with hybrid think/no-think modes, native XML tool calling, and 128K context. ⚠ not a new release (weights dropped 2026-05-19); surfaced via the trending window (08-17 repo touch = card/packaging update, no new checkpoint per GitHub changelog, latest commit 07-27).

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published (lab table is image-only)
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published
SWE-bench Multilingual—not published

The lab’s comparison table is a single image (GitHub asset public_leaderboard_en.png) and the repo ships no .eval_results/ YAMLs — no per-benchmark coding numbers in extractable form. Cells left — rather than invented.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published
AA Agentic Index—not published (agentic strength claimed in AA composite only)

Lab claims its advantage is “most visible in agentic tool use, code, and competition math”; average 42.57 across reasoning/knowledge/code/instruction-following/math/logic/agentic benchmarks vs 35.61 best for strong open-source models in the same size class — lab-reported, italicized, no independent per-benchmark split found. Tool calling is real and first-class: XML-style calls, native minicpm5 parser in SGLang, vLLM tool-parser plugin — suitable for agent loops on tiny hardware.

Reasoning (context only)

BenchmarkScoreSource
AA Intelligence Index17.9Artificial Analysis (independent — highest open-weight score at ≤1B params, +7.4 over Qwen3.5-0.8B Reasoning 10.5; beats Qwen3.5-2B Reasoning 16.3 at under half the params)
AA-Omniscience-1Artificial Analysis (independent — best in class, earned by abstaining rather than hallucinating; sub-2B peers sit at -70 to -89)
MMLU / MMLU-Pro—not published
GPQA Diamond—not published (part of AA composite only)
MATH / AIME—lab table image-only

Efficiency

  • Active params: 1.08B dense (no MoE, all active)
  • Context length: 131,072 tokens (128K native, no rope scaling needed)
  • VRAM (fp16 rough): ~2.2 GB weights (2.16 GB safetensors); q4 quant ~0.9 GB with 1–3% accuracy loss (lab-reported); ~200 ms inference on a modern laptop (AA/LinkedIn reporting)
  • Runs on: llama.cpp (GGUF), Ollama, LM Studio, MLX (4-bit, Apple Silicon), vLLM (≥0.21), SGLang (≥0.5.12, recommended for tool calling), Transformers ≥5.6 — standard LlamaForCausalLM, no custom kernels/model-code fork needed

What makes it notable

MiniCPM5-1B extends the open-weights Pareto frontier for Intelligence-vs-Params at the sub-2B scale: 17.9 AA Intelligence Index with dense 1B params beats every open model ≤2B and nearly matches Qwen3.5-2B with half the parameters. Post-training is RL + On-Policy Distillation on OpenBMB’s UltraData tiered corpus (SFT 200B deep-thinking + 200B hybrid-thinking tokens), and the same checkpoint serves both fast assistant and deliberate reasoner via enable_thinking. For local agent loops this is the strongest genuinely tiny tool-calling model: 128K context, XML tool calls with first-class SGLang/vLLM parsers, and an agent-skills/cookbook ecosystem (Cursor/Claude Code style, Ollama, desktop pet) maintained by the lab. Watch that it is text-only (no native multimodal input) and that the headline 42.57 average is lab-reported with the detail chart image-only — treat specific coding/agent numbers as unverified until third-party splits appear. For the local runner: fits in ~1–2 GB, deployable on CPU, M-series, or any small GPU.

See also

Graph View

Backlinks