MiniCPM5-2B lands on stable vLLM the day it ships, because it invented nothing
OpenBMB's small model tops its size class on a third of the tokens, and declines to answer seven out of ten questions it is asked.
OpenBMB shipped a small model this week and vLLM supported it the same day, on the stable branch. No fork, no nightly build, no custom kernels.
That was 7 September, and the serving project said why in its own announcement, and the thank-you is the interesting part: "thanks for keeping it on stock LlamaForCausalLM and opening the training data alongside the weights."
Day-zero support usually gets written up as a favour the serving project does the lab. We would put it closer to the other way round.
Restraint ships faster
MiniCPM5-2B is a 2.6B dense model on the standard Llama architecture, so there is no model code to write, no kernel to tune, no reviewer's weekend to spend. The recipe is one line, vllm serve openbmb/MiniCPM5-2B, on a single GPU. Tool calling needed one parser, merged as a single pull request into the same stable branch that made Model Runner V2 the default last month.
That is the entire integration cost, and OpenBMB paid most of it by declining to be clever (a decision nobody gets credit for).
Artificial Analysis scored the model the same morning at 15 on its Intelligence Index, the highest of any open-weights model under four billion total parameters. The next one down its size class sits at 11. Ling 3.0 Tiny beats it by a single point, with roughly three times the parameters.
Hold on to that last one.
The token bill matters more than the leaderboard place
OpenBMB's model used 19k output tokens per Intelligence Index task, joint-lowest in its comparison set. Ling spent 56k to earn that 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 thirty-seven thousand of them.
The second number is stranger, and we had to work it out
Artificial Analysis published the inputs and not the ratio. On its omniscience test the model attempts only 29 percent of questions, with an accuracy of 8 percent and a non-hallucination rate of 78 percent.
Divide the accuracy by the attempt rate. Of the questions it chooses to answer, it gets roughly 28 percent right. So it is wrong about seven times in ten when it opens its mouth.
But it mostly keeps quiet. That is probably what its score of -12 is measuring, against -53 for Qwen3.5 9B and -63 for gpt-oss-20b (both of which answer far more and pay for it). Not more knowledge. A much stricter policy on guessing.
Which is the right trade at this size, and worth saying out loud. A small model that abstains is usable inside an agent loop. A small model that confabulates is a liability with a tool call attached.
Every other score follows the same shape
Its agentic numbers lead the size class. Joint-first on a banking agent benchmark at 21 percent, against 8 percent for the next model up, and a GDPval Elo of 831 at the head of its size class.
Its knowledge numbers do not follow at all. Nine percent on Humanity's Last Exam. Nine percent on Terminal-Bench, where the Qwen model gets three times as many. Zero on CritPt. One percent on a new professional document-reasoning set.
What the bar for deployability has become
Architectural restraint plus a tool parser, and it is worth more than a leaderboard place. An Apache licence, a 131K native context, Think and No-Think from a single checkpoint, and open training data. Easy to adopt in every direction at once.
We would expect the pattern to hold. Releases on stock architectures keep landing in stable vLLM on day zero through the rest of this year, while anything with bespoke attention keeps taking weeks. A custom-architecture model getting same-day stable support is the thing that would change our minds.
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 deploys a model this small for its long context. They deploy it because it fits on the GPU they already own, and because it starts working in one command. How many models released this year can you say that about?
The open training data is the part we want other labs to copy first.
