MiniCPM5-2B open-sourced with Day-0 vLLM support, tops sub-4B open models
vllm_project · x · 2026-09-07
OpenBMB has released MiniCPM5-2B, a dense 2B model built for on-device deployment on the standard LlamaForCausalLM architecture, with native 128K context and Day-0 support in vLLM v0.21.0.
Benchmarks: It ranks #1 among open-source models under 4B parameters on the Artificial Analysis Intelligence Index with a score of 23, scores 20 on the Agentic Index, and averages 53.9 across 34 benchmarks covering coding, math, long-context, tool use, and agentic tasks.
Highlights:
- A single checkpoint supports both Think / No-Think modes, toggled via the enablethinking chat-template flag, overridable per request
- Native XML-style tool calling with vLLM's minicpm5 parser
- Training data, recipes, and the RL stack are open-sourced alongside the weights
- Trivial to deploy: vllm serve openbmb/MiniCPM5-2B on a single GPU; smaller cards can shrink --max-model-len to free KV cache
Related event: OpenBMB Releases MiniCPM5-2B, Tops Sub-4B Open Model Intelligence Index(4 posts)→
More from coding & agent
- YC Teases Live Stream of Harnesses Deep Dive: 30% to 95% on ARC-AGI — ycombinator · 2026-09-07
- YC Harnesses Deep Dive: Same Weights Jump From 30% to 95% on ARC-AGI — ycombinator · 2026-09-07
- Even seasoned devs now look at vibe-coded apps and 'have no idea how that's made' — jxnlco · 2026-09-07
- Why RL environments work now (and couldn't in 2016): TRL + OpenEnv explained — SergioPaniego · 2026-09-07
- HF engineer's DIY continual learning bench: SFT works, SDPO doesn't yet — ben_burtenshaw · 2026-09-07
- Autonomous launches $149 Harness, a desk device to orchestrate every coding agent — dee_hw · 2026-09-07