Intern-Decision-2B vs MiniCPM5-2B: Two 2B Checkpoints, Twenty Days Apart, Opposite Ways to Spend the Parameters

internlm/Intern-Decision-2B and openbmb/MiniCPM5-2B are the same size, released twenty days apart, licensed the same way, and built for nothing resembling the same job. InternLM's checkpoint is 2,213,241,664 parameters of decision scorer: it takes a state and typed questions, runs one forward pass, and returns calibrated probabilities without generating a token, refusing any input past 8,192 tokens. OpenBMB's is 2,516,756,480 parameters of dense generalist: a 42-layer Llama derived from the MiniCPM5 recipe, taking 131,072 tokens of context, writing code, calling tools and doing mathematics, with an Artificial Analysis Intelligence Index of 12.5 and 726,518 downloads last month. They cost the same to download and they buy entirely different things.

The interesting part of this pairing is not the benchmark gap — there is barely an overlapping benchmark to compare. It is what a 2.2-billion and a 2.5-billion parameter budget can each be spent on, and what the choice tells you about which one is finished. MiniCPM5-2B is the second model in its series, following MiniCPM5-1B from May, and it arrived with a full release apparatus. Intern-Decision-2B is the middle of three sizes InternLM pushed to Hugging Face at 05:36 UTC on September 26, 2026 with no announcement, and its release apparatus — a training codebase, a hash-verified evaluation bundle, a 96-case calibration benchmark — only became public this morning at 08:58 UTC.

Where the two budgets went

Both models are fine-tunes of somebody else's architecture, and both occupy roughly 2.5 billion parameters. That is where the resemblance ends.

MiniCPM5-2B is a dense Transformer with 42 layers, hidden size 2,048, 16 attention heads, an intermediate size of 6,144, a 130,560-token vocabulary and a 131,072-token context, trained on a mixture of open corpora OpenBMB published alongside it: UltraX for web pre-training, UltraData-Code with L0–L3 tiered code management, UltraData-Math, and 500,000 agent SFT samples with over 80,000 RL samples in UltraData-RL-2609. Its post-training runs SFT, then specialised RL teachers in mathematics, code and agentic tasks, then on-policy distillation back into one model. It is a general-purpose model whose entire parameter budget is exposed as capability you can prompt.

Intern-Decision-2B is a Qwen3_5ForConditionalGeneration with 24 layers — a repeating pattern of three linear-attention layers to one full-attention layer — hidden size 2,048, 8 query heads against 2 key-value heads, head dimension 256, a retained multi-token-prediction layer and a 262,144-position embedding ceiling. It ships with a 612.5 MB vision tower and a 50.3 MB projector. Almost none of that budget is available to you as a prompt: the model answers a fixed schema of questions, and its architecture is the delivery mechanism for a softmax, not a text generator.

One detail from InternLM's newly published repository is worth pulling out, because it reframes the multimodal tag on the model card. The decision tuning "fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector." Both the 2B and the 4B decision checkpoints carry a vision shard of exactly 612,517,440 bytes — the same size in both — which is consistent with those components being inherited from the Qwen base rather than trained. The image path is real and it is usable, but it is not what InternLM's decision tuning spent its effort on. If your workload is text-only decision scoring, roughly 660 MB of that 4.46 GB download is doing nothing for you.

Scoreboard

• Parameters — Intern-Decision-2B 2,213,241,664 in BF16 plus circa 660 MB of vision tower and projector, about 4.46 GB of repository; MiniCPM5-2B 2,516,756,480 in BF16, one 5.03 GB safetensors file.

• Context — 8,192 tokens for Intern-Decision-2B, rejected without truncation above that; 131,072 tokens for MiniCPM5-2B.

• Output — a calibrated distribution plus an argmax per field, no text at all; generated text, token by token.

• Training — decision tuning of a Qwen3.5-2B backbone with the vision tower frozen, training data undisclosed; a full pre-train, mid-train and 400-billion-token deep-thinking SFT pass followed by RL and on-policy distillation, with the corpora published alongside.

• Published evidence — a seven-suite vendor table averaging 84.68 with Brier 0.437 and ECE 0.100, unreproduced by any third party, one like and zero downloads; an OpenBMB card averaging 53.9 within its own comparison set and an independent Artificial Analysis Intelligence Index of 12.5, with 726,518 downloads in the last month.

• Licence — Apache-2.0 with the upstream Qwen terms preserved as LICENSE-QWEN, and no licence stated at all on the code repository; Apache-2.0 throughout, code included.

What each one is actually better at, and it is not close

The comparison is asymmetric to the point of being a category error, so it is more useful to name the jobs.

MiniCPM5-2B wins everything that involves producing an answer rather than selecting one. It writes code; Intern-Decision-2B has no code path. It handles a 131,072-token context; Intern-Decision-2B stops at 8,192 and fails loudly. It calls tools, with a BFCL v4 score of 66.6 on OpenBMB's own table, and it does mathematics, with MATH-500 at 94.6 and GPQA-Diamond at 70.2 — figures from the vendor's card, with the GPQA row carrying a footnote the card itself supplies. It supports 70.8 on MMLU-Pro and 84.7 on MMLU-Redux. It runs in llama.cpp and MLX, ships in GGUF, GPTQ and DSpark variants, and has a demo Space on the Hub that is public, unlike the 401 that Intern-Decision's card points at.

Intern-Decision-2B wins exactly two things, and they are the reason it exists. First, a closed-set decision with a schema you write returns a probability you can threshold, in a single forward pass, in about 33 milliseconds, with no parser and no sampling — a shape MiniCPM5-2B cannot produce except by generating text and hoping the number inside it is well-behaved. Second, it is a reproducible artefact: the repository now carries the training objective, the seven accuracy suites with SHA-256 hashes and per-suite row counts, the scoring code, the temperature-fitting and replay scripts, and a 96-case distribution calibration benchmark with its own generator and offline scorer. InternLM's evaluation guide notes that the released presets are bound to the XTuner backend and that HF results should be reported as a different execution setting — which is exactly the kind of caveat a reproduction kit exists to make checkable.

Who should download which

If you have a task that involves reading something long, writing something, or answering a question whose options you did not enumerate in advance, MiniCPM5-2B is the model and there is no decision to make. It is a finished 2B-class generalist with a real ecosystem, open data behind it and an independent evaluation listing, and it costs nothing to try — GGUF and MLX builds are already on the Hub.

If you have a task that is genuinely a choice among known answers — routing a ticket, classifying a support email, labelling a candidate, scoring an intent on an ordinal scale — then a generation model is the wrong instrument, and Intern-Decision-2B is the more interesting of the two even though almost nobody has run it. A probability you can threshold is cheaper to operate, easier to audit and impossible to hallucinate, and the fact that the checkpoint arrived with a reproduction kit rather than a leaderboard post makes it the better-documented of the two on the axis that matters when you have to defend the choice.

Neither model is on OrcaRouter. Our model page for Intern-Decision-2B returns 404 and there is no MiniCPM5-2B route; nothing here is an availability claim, and both mean running your own inference. Where the practical alternative exists is in the hosted decision models: TypeSafe's Jev 1.13, the model InternLM benchmarked its whole family against, is on the catalogue at $0.042 per million input tokens with a 65K context and a P50 time-to-first-token of 178 ms. And if what you actually need is a generalist in the same size class as MiniCPM5-2B without the operational work, our smallest hosted Qwen route is several times the size but is one key among 200-plus models, at provider list price passed through with no markup and automatic failover behind it.

The one number that decides it

It is not 84.68 against 53.9. Those two averages come from different suites, measured by different labs, and in one case by a lab that has never had its numbers checked. The number that decides it is 8,192. If your state fits inside eight thousand tokens and your answer is already a list, Intern-Decision-2B is a precise instrument with a reproduction kit and a calibration anomaly you should be aware of — its expected calibration error of 0.100 is the worst of the three sizes InternLM published, against 0.066 for the model half its size, so fit your own temperature before trusting a confidence value. If your state does not fit, the question is over before the benchmarks start, and MiniCPM5-2B is the answer.