---
format: "aidr-story-markdown/v1"
id: "556d21906ffc395fe93154c20994f25b994333c047934d1d6d145f36e0e745f6"
canonical_url: "https://aidr.today/556d2190?lang=en"
title: "OpenBMB MiniCPM5-2B Tops Open Models Under 4B Parameters With Intelligence Index 15"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-09-07T16:44:39.000Z"
category: "Releases"
topics: ["open-source","llm","agent","vllm","inference"]
source_urls: ["https://huggingnews.com/ai/openbmb-minicpm5-2b-tops-open-models-under-4b-parameters-with-intelligen-ed753446","https://x.com/OpenBMB/status/2096970974247956501","https://x.com/ArtificialAnlys/status/2096955784592797998","https://x.com/vllm_project/status/2096956274428772654","https://x.com/sgl_project/status/2096973451307401577","https://x.com/OpenBMB/status/2096953469961957775","https://x.com/ArtificialAnlys/status/2096955804033401328"]
summary: "OpenBMB's latest 2.6 billion parameter dense reasoning model is available as open source under the Apache 2.0 license. The software outperformed the 3 billion parameter Granite 4.2 by 4 points on the Artificial Analysis Intelligence Index v4.2, where it matched the performance of Qwen3.5 9B. MiniCPM5-2B supports a 131,000 token context window, accepts text-only input, and offers a non-hallucination rate of 78% on the AA-Omniscience test by abstaining from 71% of questions. The release includes the full data pipeline—covering web, code, and RL components—and training recipes. Serving support launched on day one via vLLM, which includes tool calling support, and SGLang, which showed decode speeds over 250 tokens per second per user on Nvidia RTX 5090 cards for coding tasks. The model scored 20 on the Agentic Index, 10 times higher than the score of 2 recorded for Mistral 3 3B or LFM2.5-2.6B."
---

# OpenBMB MiniCPM5\-2B Tops Open Models Under 4B Parameters With Intelligence Index 15

> [Open the canonical story](<https://aidr.today/556d2190?lang=en>)

**Published:** 2026-09-07T16:44:39.000Z
**Category:** Releases
**Topics:** open\-source, llm, agent, vllm, inference

## Summary

OpenBMB's latest 2\.6 billion parameter dense reasoning model is available as open source under the Apache 2\.0 license\. The software outperformed the 3 billion parameter Granite 4\.2 by 4 points on the Artificial Analysis Intelligence Index v4\.2, where it matched the performance of Qwen3\.5 9B\. MiniCPM5\-2B supports a 131,000 token context window, accepts text\-only input, and offers a non\-hallucination rate of 78% on the AA\-Omniscience test by abstaining from 71% of questions\. The release includes the full data pipeline—covering web, code, and RL components—and training recipes\. Serving support launched on day one via vLLM, which includes tool calling support, and SGLang, which showed decode speeds over 250 tokens per second per user on Nvidia RTX 5090 cards for coding tasks\. The model scored 20 on the Agentic Index, 10 times higher than the score of 2 recorded for Mistral 3 3B or LFM2\.5\-2\.6B\.

## Sources

- [Story source](<https://huggingnews.com/ai/openbmb-minicpm5-2b-tops-open-models-under-4b-parameters-with-intelligen-ed753446>)
- [Story source](<https://x.com/OpenBMB/status/2096970974247956501>)
- [Supporting source](<https://x.com/ArtificialAnlys/status/2096955784592797998>)
- [Supporting source](<https://x.com/vllm_project/status/2096956274428772654>)
- [Supporting source](<https://x.com/sgl_project/status/2096973451307401577>)
- [Story source](<https://x.com/OpenBMB/status/2096953469961957775>)
- [Supporting source](<https://x.com/ArtificialAnlys/status/2096955804033401328>)