Intern-Decision-0.8B vs Gemma 4 12B: A Scoring Head Against a Multimodal Generalist

The cheapest way to see what Intern-Decision-0.8B is — and what it is not — is to put it next to Gemma 4 12B, and then notice that the comparison is almost entirely about what each model does with its output layer. Gemma 4 12B, DeepMind's 11.95-billion-parameter encoder-free multimodal model released on 3 June 2026 under Apache 2.0, is a generative generalist: you give it text, images, audio or video, and it writes an answer. Intern-Decision-0.8B, uploaded quietly by InternLM to Hugging Face on 26 September 2026 with no announcement, is a fine-tune of the much smaller Qwen3.5-0.8B that produces no text at all — it takes a state, a schema of named questions and up to eight images, and returns a calibrated probability distribution for each question after a single forward pass. One writes. The other scores. Everything below follows from that.

That makes this a strange matchup to frame as a contest, and worth saying plainly before the spec rows: these two are not substitutes, and anyone choosing between them has already mis-stated the problem. Gemma 4 12B answers the question "what should the response be." Intern-Decision-0.8B answers the question "which of these sixteen options is it, and how confident are you" — and answers it without a decoder loop, which is where the latency and cost differences come from. The useful comparison is therefore about how you would wire each into one workflow, not about which wins.

The single largest difference is the output side of the bill

Gemma 4 12B is billed like any generative model: input tokens and output tokens, both. When you ask it for structured JSON, every one of those tokens is generated, sampled and billed, and the longer and more nested your schema, the more of them there are. That is not a criticism of the model — it is what a decoder does — but it is the line item that decision heads exist to remove. Intern-Decision-0.8B has no output tokens. Its card is explicit about why: the inference path never calls generate()/, reading logits at the positions immediately before each <decision>/ placeholder in a pre-rendered skeleton and taking a softmax over only the permitted candidate symbols for that field. The usage counter called output_tokens/ counts scored fields, not text.

The consequence compounds at scale. A pipeline that labels ten thousand records with Gemma 4 12B pays for ten thousand JSON documents; the same pipeline on Intern-Decision-0.8B pays for the prompts and nothing else. Whether the absolute number is large depends entirely on your volume and schema, and anyone with a per-token budget can work it out in a spreadsheet in ten minutes. But the direction is not in dispute, and it is the reason a category of small decision models has appeared at all.

Latency, and why the parameter gap is misleading

• Throughput model — Intern-Decision-0.8B: one forward pass, no decoding loop. Gemma 4 12B: autoregressive generation, one token at a time.

• Measured latency — Intern-Decision-0.8B: 33.98 ms mean / 37.50 ms p95 on a single RTX 4090 via the local Hugging Face path, per InternLM's own measurement. Gemma 4 12B: no comparable single-pass figure exists, because there is no single pass to measure.

• Footprint — Intern-Decision-0.8B: about 1.73 GB of repository storage, 852,985,920 parameters across a 1.50 GB language shard, a 176 MB vision shard and a 25 MB projector. Gemma 4 12B: Google's launch materials put it at 16 GB of VRAM or unified memory.

• Output determinism — Intern-Decision-0.8B: argmax over candidate logits, so the same input yields the same label every time. Gemma 4 12B: sampling unless you pin the temperature to zero, and even then the label arrives as text you must parse.

The 14x parameter gap between these two models does not translate into anything like a 14x runtime gap on a decision task, because the work is different in kind. Intern-Decision-0.8B spends its one pass reading the prompt — the state, the schema, the option descriptions — and then stops. Gemma 4 12B spends that same prompt-read and then adds every token of the answer on top. For a sixteen-field schema with descriptive option labels, the generation leg is not a rounding error; it is most of the wall clock. InternLM's own table makes the point accidentally: its 2B sibling posts a 33.28 ms mean, marginally faster than the 0.8B's 33.98 ms. When prompt processing dominates, parameter count stops being the lever. img src="2.png">

Where Gemma 4 12B is simply in a different category

It would be dishonest to leave the spec sheet at the decision-task framing, because outside that framing Gemma 4 12B is not merely ahead — it is playing a different sport.

Intern-Decision-0.8B accepts text plus up to eight images, and that is the whole of its input surface. Its default input ceiling is 8,192 tokens, rejected rather than truncated, which is well below the 262,144 maximum position embedding its config advertises — the ceiling, not the architecture, is the practical window. Gemma 4 12B takes text, image, audio and video natively through an encoder-free design that projects raw patches and waveforms straight into the embedding space, holds a 256K context window, and returns text. Google's published card scores it at 77.2% on MMLU Pro, 77.5% on AIME 2026 without tools, 72.0% on LiveCodeBench v6, 78.8% on GPQA Diamond, 69.1% on MMMU Pro and 79.7% on MATH-Vision. Those are vendor figures from Google's own evaluation card, and unlike Intern-Decision-0.8B's table they have had a year of third-party use behind them, because Gemma 4 shipped into an ecosystem — LM Studio, Ollama, llama.cpp, MLX, vLLM, SGLang, Unsloth — on day one.

The releases also look nothing alike, and the contrast is instructive. Gemma 4 12B had a launch blog, a technical report on arXiv, a five-model family page, an evaluation card, and a download link on the day it appeared. Intern-Decision-0.8B had a model card with a GitHub URL that returns 404 and a demo Space that returns 401, and it appeared within forty seconds of two larger siblings that are equally undocumented. One of these is a product. The other is a checkpoint someone decided to put on the internet.

Both are Apache 2.0, and that is not a trivial shared row

The one dimension where the two genuinely meet is licensing, and it is worth a paragraph because it is where the decision changes for commercial users. Both are Apache 2.0. Gemma 4 12B ships under the Gemma 4 licence terms hosted at Google, which the model card identifies as Apache 2.0; Intern-Decision-0.8B carries Apache-2.0 alongside a second LICENSE-QWEN/ file, which is the correct handling for a model derived from Qwen weights and a signal that InternLM did the paperwork even for a release it did not announce.

That distinguishes both from Liquid AI's LFM2.5 family, whose LFM Open License v1.0 conditions commercial rights on your legal entity staying under a $10 million annual revenue threshold — a real constraint that a decision-model buyer comparing options should read before assuming "open weights" means the same thing everywhere. If your use case is commercial and your revenue clears that line, Apache 2.0 and the LFM licence are not interchangeable, and neither Gemma 4 12B nor Intern-Decision-0.8B carries the restriction.

The honest ledger on Intern-Decision-0.8B's numbers

Every performance figure for Intern-Decision-0.8B is vendor-reported and unreproduced. That is not a formality — it is the current state of the evidence, and it should govern how much weight the table carries. InternLM selected the benchmarks, selected the competitors, ran the evaluations and published the results, and no independent run exists on any public leaderboard. An Artificial Analysis lookup for the model returns nothing at all. img src="3.png">

With that label attached, the shape of the result is still informative. The 0.8B posts 77.35 on TypeSafe's Typed Decision benchmark against 73.35 for Jev 1.13 — the model the benchmark is named after — and 94.52 on ToolACE against 91.29. It averages 79.38 across the suite, with its own 2B and 4B siblings at 84.68 and 90.02. The column that should give a buyer pause is WildJailBreak, where it scores 64.48 against Jev's 96.29: a refusal-robustness gap of that size is a property of the fine-tune, and whether it comes from the Qwen3.5-0.8B base, the decision-tuning objective or the parameter count is not documented anywhere. Its calibration profile is the more interesting reading — a Brier of 0.530 and an expected calibration error of 0.066, against Jev's 0.358 and 0.095 — meaning it is less accurate overall but its stated confidences track its accuracy more closely. For a threshold-based workflow that distinction matters more than raw accuracy, and it is the kind of thing InternLM had no incentive to publish.

Against that, Gemma 4 12B's evaluation card is also vendor-reported — Google ran those benchmarks too — but it has been in the world since June, has been deployed through every major local runtime, and any discrepancy would have surfaced. The evidentiary gap between "unreproduced for a week" and "unreproduced for four months and nobody has contradicted it" is the real difference between these two columns.

How you would actually combine them

The pattern that makes sense is not a choice. A decision head handles the structured calls — routing, triage, classification, scoring against a rubric — where the answer set is closed, latency matters and output tokens are pure overhead. A generalist handles everything that needs prose, code, synthesis or long context: the escalation summary, the explanation of why the ticket went to billing, the draft that a human edits.

If you are working out where the boundary sits, the cheapest experiment is to put both halves behind one endpoint. OrcaRouter routes 200+ models through a single OpenAI-compatible API with automatic failover and provider list price passed through at zero markup, which means testing a hosted decision model and a hosted generalist in the same script costs one key and one afternoon. Worth being precise about one boundary here: Intern-Decision-0.8B is not one of ours — it is an Apache-2.0 checkpoint you download and run yourself, and the whole reason to pick a 1.73 GB model is that it lives where your data already does. The generative half of the pattern is the part that is one line away. As for Gemma 4 12B specifically, it is not on our catalogue either; the Gemma 4 sizes we do route are 31B and 26B A4B, and the 12B is a local download. img src="4.png">

The verdict, then, is not a winner. Intern-Decision-0.8B is a cheap, fast, deterministic scoring head from a lab that has not yet explained it, with a benchmark table nobody has reproduced and a refusal-robustness column that should be tested before it goes near untrusted input. Gemma 4 12B is a mature open-weights multimodal generalist with a published card, a technical report and fourteen times the parameters, and it is the wrong tool for a sixteen-field classification call — not because it is worse, but because generating an answer to a question whose answer set you already wrote down is an expensive way to get a label.