Intern-Decision-2B vs Gemma 4 12B: An 8,192-Token Ceiling Against a 256K Window

Suppose the thing you need judged is a forty-page insurance claim file. internlm/Intern-Decision-2B will refuse it. Not truncate it, not summarise it — refuse it, because the bundled inference engine declares DecisionEngine(max_length=8192) and the model card says in plain language that oversize inputs are "rejected without truncation." google/gemma-4-12B-it — Gemma 4 12B Unified — will take the same document, plus the scanned photographs stapled to it, plus the recorded phone call, and write you an answer. That contrast is the whole comparison, and it has almost nothing to do with the parameter counts — 2,213,241,664 against 11,959,730,224 — that a spec-sheet reading would put first.

Intern-Decision-2B is the middle of three decision checkpoints InternLM uploaded to Hugging Face on September 26, 2026 without an announcement, fine-tuned from Qwen/Qwen3.5-2B and licensed Apache-2.0 with the Qwen terms preserved alongside. Gemma 4 12B Unified is Google DeepMind's encoder-free multimodal model from May 2026, an any-to-any system taking text, image, audio and video and generating text across a 256,000-token window in over 140 languages. One of these is a scorer. The other is a generalist that happens to be able to score badly if you prompt it carefully. Deciding between them is less a benchmark question than a question about what shape your answer already is.

Where the two are not actually comparable

It is worth being blunt about this before the numbers, because the matchup invites a false symmetry.

Intern-Decision-2B produces no tokens. It takes a state, one to sixteen named questions with up to 62 options each, and optionally up to eight images; renders an assistant JSON skeleton with one <decision> placeholder per field; runs a single causal forward pass; reads logits at the position immediately before each placeholder; softmaxes over only that field's legal symbols; and applies a fitted temperature of 2.100509348278. The output is a calibrated distribution and an argmax per field. The card states the negative outright: "This API performs structured candidate scoring. It does not call generate() or sample free-form text."

Gemma 4 12B generates. Everything it does — reasoning with configurable thinking, function calling, OCR, chart comprehension, speech recognition, 140-language coverage — runs through a decoding loop over a 262,144-entry vocabulary. Ask it for a routing decision with probabilities and you will get prose containing a number, and the number will be whatever the sampler produced, not a softmax over your option set.

So the practical question is not "which is more accurate." It is "is my answer already a closed set?" If it is, the 12B is a wasteful way to pick from a list. If it is not, Intern-Decision-2B cannot help you at any accuracy at all.

The input ceiling is the sharpest line between them

Here is the dimension to weigh above all others, because it produces hard failures rather than degraded ones.

• Input length — Intern-Decision-2B accepts 8,192 tokens and rejects anything longer, loudly; Gemma 4 12B accepts 256,000 tokens and, on Google's own card, scores 43.4% on MRCR v2 eight-needle at 128k.

• Images — up to eight for Intern-Decision-2B, and they count against the same 8,192-token budget; variable aspect ratio and resolution for Gemma 4 12B, with interleaved text and image input in any order.

• Modalities in — text and image for one; text, image, audio and video for the other.

• Language — no multilingual claim is made anywhere in the Intern-Decision documentation; Gemma 4 covers 140+ pre-trained languages with an evaluated multilingual score of 83.4 on MMMLU for the 12B.

• Output — a typed JSON answer distribution for one; generated text for the other.

The 8,192 limit is not arbitrary, and that is what makes it hard to work around. The decision objective reads a logit at a fixed position per field, so the prompt must contain the state, the schema and the full skeleton in one pass; there is no chunking strategy that preserves the contract. A ticket, an email, a short JSON state, a screenshot or two — that is the design centre. A document that needs retrieval is a document this model is not for.

What each one costs you, honestly counted

Intern-Decision-2B has no hosted endpoint and no provider. OrcaRouter's model page for it returns a 404, and nothing in this article is an availability claim. Calling it means downloading roughly 4.46 GB of repository — a 3.76 GB language shard, a 612.5 MB vision tower, a 50.3 MB projector — and running inference.py beside the weights in a Python 3.12 environment with torch 2.9.1 and transformers 5.14.1, on a GPU you are paying for whether or not it is scoring decisions.

That is not a disadvantage in every case, and the arithmetic is worth doing rather than assuming. At 33.28 ms mean per request on a single RTX 4090, in InternLM's own measurement, a locally hosted Intern-Decision-2B bills nothing per decision. A hosted generalist bills per token, including for the tokens it spends thinking before it answers. At volume, a scorer you already own is the cheaper instrument — the cost is the GPU and the engineering, not the call.

Gemma 4 12B Unified is likewise not on our catalogue; if you want it you fetch 11.96 billion parameters of BF16 and serve it yourself. Where our catalogue does have real Gemma 4 routes is the other two sizes, and they are cheap: Gemma 4 26B A4B at $0.06 per million input tokens and $0.33 per million output, and Gemma 4 31B at $0.13 and $0.38, both listed with 262,144-token contexts and a P50 time-to-first-token the catalogue shows at 1.73 seconds. If your problem turns out to be general reasoning rather than a closed-set decision, that is the thing to compare against a locally operated scorer — two more models on one key, provider list price passed through with no markup, and automatic failover so that a model you are still evaluating never becomes a single point of failure in a production path.

Scoreboard, with the caveats attached

• Parameters — Intern-Decision-2B 2,213,241,664 in BF16 plus a vision tower and projector; Gemma 4 12B Unified 11,959,730,224 in BF16.

• Context — 8,192 tokens, rejected above, versus 256,000 tokens.

• Output — calibrated probabilities and an argmax, no text; generated text, one token at a time.

• Modality — text plus up to eight images versus text, image, audio and video in, text out.

• Published evidence — a seven-suite vendor table averaging 84.68 with a Brier score of 0.437 and ECE 0.100, unreproduced by anyone outside the lab; a Google evaluation card with MMLU Pro 77.2%, GPQA Diamond 78.8%, AIME 2026 without tools 77.5% and LiveCodeBench v6 72.0, plus an independent listing at an Artificial Analysis Intelligence Index of 14.2.

• Adoption — one like and zero downloads on the 2B checkpoint against a family that has been in circulation since May with quantisations, fine-tunes and derivatives numbering in the hundreds.

Read the evidence rows asymmetrically, because that is what they are. Google's numbers have been independently indexed and reproduced by third parties; InternLM's have not been run by anyone outside the lab, and the calibration figure in particular should be treated as a well-documented intention until it meets a foreign test set. The honest summary is not that the vendor table is wrong. It is that the two columns do not carry the same evidentiary weight, and a comparison that prints 84.68 beside 77.2 without saying so is misleading its reader.

What to do with this

Choose Intern-Decision-2B when the decision is genuinely closed, the state fits comfortably inside eight thousand tokens, you want a probability you can threshold rather than a paragraph you have to parse, and you have the hardware and the appetite to operate a checkpoint nobody supports yet. The middle size specifically carries the weakest published calibration of the three InternLM shipped — ECE 0.100 against 0.066 for the model half its size and 0.065 for the one nearly twice it — so if you are picking within this family, the 2B is the one to fit your own temperature on, not the one to trust out of the box.

Choose Gemma 4 12B — or more practically, one of the two Gemma 4 sizes we actually route — when the input is longer than a page, when audio or video or a language other than English is involved, when the answer must be explained rather than merely selected, or when you do not want to own the inference stack. The 12B is the smallest of the Gemma 4 models with native audio; the 26B A4B activates about four billion parameters per token and is the cheapest way into the family.

And if what you actually have is a scoring problem with a long document attached, neither of these is the right instrument: that is a retrieval problem wearing a comparison article's clothes.