MiniCPM5-2B lands on stable vLLM the day it ships, because it invented nothing

OpenBMB released MiniCPM5-2B on September 7th, and vLLM supported it the same day on the stable branch. No fork, no nightly, no custom kernels. The vLLM project said why in its own announcement, and the thanks is the interesting part: "thanks for keeping it on stock LlamaForCausalLM and opening the training data alongside the weights".

vLLM@vllm_project

🤝 Day-0 support for MiniCPM5-2B on stable vLLM.

⚡ Dense 2.6B model with 131K native context
🧠 Think / No-Think from the same checkpoint
🔧 Tool Calling support via vLLM’s minicpm5 parser

Congrats @OpenBMB on the release, and thanks for keeping it on stock LlamaForCausalLM and opening the training data alongside the weights! 🙌

🔗 t.co/eZyYwNAPcM

Day-0 support is usually written up as a favour the serving project does the lab. We would put it closer to the reverse. A 2.6B dense model on the standard Llama architecture needs no model code written, no kernel tuned, no reviewer's weekend. The vLLM recipe for it is one line, vllm serve openbmb/MiniCPM5-2B --port 8000, on a single GPU at TP=1. Tool calling needed one parser, minicpm5, merged as PR #43175. That is the whole integration cost, and OpenBMB paid most of it by declining to be clever. Restraint ships faster.

Artificial Analysis scored it the same morning at 15 on Intelligence Index v4.2, which it calls the highest of any open-weights model under 4B total parameters. Granite 4.2 3B, the next one down, sits at 11. Ling 3.0 Tiny beats it by a point with roughly three times the parameters. Hold that thought.

Artificial Analysis@ArtificialAnlys

OpenBMB's MiniCPM5-2B scores 15 on the Artificial Analysis Intelligence Index v4.2, the highest of any open weights model under 4B total parameters

OpenBMB (@OpenBMB) is the open-source AI group behind the MiniCPM series of efficient small models. MiniCPM5-2B is a 2.6B parameter dense reasoning model with text input and output, released under Apache 2.0.

Scoring 15 on the Intelligence Index, MiniCPM5-2B sits one point behind Ling 3.0 Tiny (16), which has ~3x the total parameters. Among open weights models under 4B total parameters, the next best score is Granite 4.2 3B (11).

Key results:

➤ The highest Intelligence Index of any open weights model under 4B total parameters, setting a new Pareto-optimal point on Intelligence vs. Total Parameters: Its score of 15 is 4 points clear of Granite 4.2 3B (11). With 2.6B total parameters, it is 1 point ahead of Qwen3.5 4B (Reasoning, 14, estimated) with 44% fewer parameters, and level with Qwen3.5 9B (Reasoning, 15, estimated) at roughly 4x its size. As a dense model, its size advantage is in memory footprint rather than active-parameter compute.

➤ Strong agentic performance at this size: Its GDPval-AA v2 Elo of 831 leads <4B models, and on τ³-Banking it is joint-first with Ling 3.0 Tiny at 21%, compared to 8% for the next best model, Granite 4.2 8B. On AA-Briefcase, it placed second among the measured models in the comparison set with an Elo of 438, above Granite 4.2 8B (324) and just below Ling 3.0 Tiny (485).

➤ Knowledge, coding and long context are where it gives ground: MiniCPM5-2B places 7th in the set on Humanity's Last Exam (9%, behind Gemma 4 12B (Reasoning) at 16%), 8th on Terminal-Bench v2.1 (9%, behind Qwen3.5 9B (Reasoning) at 29%) and scores 0% on CritPt. On SciCode it is second of the five measured models at 26%, behind Granite 4.2 8B (31%). On AA-LCR v1.1 it scores 59%, 5th in the set, one point behind Ling 3.0 Tiny (60%). On GDP.pdf, our new professional document reasoning evaluation, it passes 1% of tasks outright, behind gpt-oss-20b (high) at 2%.

➤ Its AA-Omniscience score of -12 is earned by abstaining from answering rather than accuracy: MiniCPM5-2B attempts only 29% of AA-Omniscience questions, giving it a Non-Hallucination Rate of 78%. Its accuracy of 8% is a point below Ling 3.0 Tiny (9%) and half that of Qwen3.5 9B (Reasoning, 16%). Peers that attempt far more questions are penalized heavily, with Qwen3.5 9B (Reasoning) at -53 and gpt-oss-20b (high) at -63.

➤ It is token-efficient for a reasoning model: MiniCPM5-2B used 19k output tokens per Intelligence Index task, joint-lowest in the comparison model set with Granite 4.2 3B (19k). Ling 3.0 Tiny spends 56k, roughly 3x as many, for 1 more index point.

Additional model details:

➤ Parameters: 2.6B (dense)

➤ Context window: 131k tokens

➤ Input modalities: Text only

➤ License: Apache 2.0

But the leaderboard line is the least useful number in that post. Which two should you look at instead, if you plan to actually run this thing?

The first is the token bill. MiniCPM5-2B used 19k output tokens per Intelligence Index task, joint-lowest in its comparison set. Ling 3.0 Tiny spent 56k to score its extra point. So the model one place above it costs about three times as much to run per task, and if you are paying for reasoning tokens rather than reading benchmark tables, one index point is not obviously worth 37,000 tokens.

The second is stranger, and we had to do the arithmetic ourselves (Artificial Analysis published the inputs and not the ratio). On AA-Omniscience the model attempts only 29% of questions and scores 8% accuracy, with a non-hallucination rate of 78%. Divide the second by the first: of the questions it chooses to answer, it gets roughly 28% right. It is wrong about seven times in ten when it opens its mouth. And it mostly keeps quiet. That is probably what the -12 Omniscience score is measuring, against Qwen3.5 9B's -53 and gpt-oss-20b's -63: not more knowledge, but a much stricter policy on guessing.

Which is the right trade for a 2.6B model, and worth saying out loud. A small model that abstains is usable in an agent loop. A small model that confabulates is a liability with a tool call attached. Its agentic scores follow the same shape: 21% on τ³-Banking against 8% for Granite 4.2 8B, an 831 GDPval-AA v2 Elo leading the under-4B field. Its knowledge scores do not follow it at all: 9% on Humanity's Last Exam, 9% on Terminal-Bench v2.1 where Qwen3.5 9B gets 29%, 0% on CritPt, 1% on AA's new GDP.pdf document-reasoning set.

Our read is that the deployability bar in 2026 has quietly become architectural restraint plus a tool parser, and it is worth more than a leaderboard place. Apache 2.0, 131K native context, Think and No-Think from one checkpoint via enable_thinking, and open training data — the release is easy to adopt in every direction at once. We would expect the pattern to hold: releases that ship stock architectures will keep landing in stable vLLM on day zero through the rest of the year, while anything with bespoke attention keeps taking weeks. A custom-architecture model getting same-day stable support would be the thing that changes our mind.

The counter-argument is that restraint has a ceiling. You cannot build sparse attention, a novel router or a linear-attention hybrid out of stock LlamaForCausalLM, and the models winning on long context this year are all doing exactly that. True. But almost nobody is deploying a 2.6B model for its 131K context. They deploy it because it fits on the GPU they already own (the recipe suggests dropping --max-model-len to 8192 on a small card), and because it starts working in one command.

The open training data is the part we want other labs to copy first.