Glimpse vs Intern Decision 4B: Which of These Two Actually Answers Your Question

The two models in this comparison share a modality and almost nothing else. E​lorian G​limpse, released 16 October 2026, is a visual-reasoning preview that can return to a window inside a video it has already watched — in the published trace, nine-tenths of a second re-scanned at around 14 frames per second after a first pass at roughly 8 — and then writes an answer in prose. Intern-Decision-4B, uploaded by Shanghai AI Laboratory's I​nternLM team on 26 September 2026 with no announcement at all, does not write anything. It takes a shared state, a schema of named questions and optionally some images, and returns a calibrated probability distribution across those named answers in a single forward pass. It never calls generate(). There is no chain of thought, no sentence, no token-by-token output of any kind.

Which means the first question in this comparison is not which model is better. It is whether your problem is a reading problem or a decision problem, and most teams discover they have been treating a decision problem as a reading problem because reading was the only instrument available.

A model that returns a number instead of a paragraph

The Intern-Decision family arrived quietly on the morning of 26 September 2026 — three sizes, 0.8B, 2B and 4B, uploaded within forty seconds of one another, which is the signature of a script rather than a release process. There was no blog post, no model card essay, no launch thread. The demo Space returned 401 at the time of writing, the collection returned 401, the GitHub organisation returned 404, and the largest member of the family had accumulated zero downloads and one like. Everything below is read off the released configuration and repository rather than from a vendor telling us what it does.

What the configuration says is specific. Intern-Decision-4B is a 4.54-billion-parameter fine-tune of Q​wen 3.5 4B, and its interface is a decision engine rather than a chat completion. You hand it a shared state — the context that all the questions share — plus a schema naming a set of questions, plus up to eight images. It returns a distribution over a fixed vocabulary of single-token symbols: A through Z, a through z, and 0 through 9. Sixty-two symbols, which is exactly why the option cap is sixty-two, and not a design choice anyone arrived at by thinking about how many answers a human wants. Three question types are supported: a choice between named options, a score on a scale, and a yes/no-shaped decision.

That is the whole product. The value of it is not accuracy in the abstract. It is that the output is a probability you can put a threshold on and audit afterwards, produced deterministically in one forward pass, rather than a paragraph whose confidence you have to infer from its tone.

The interface is the real difference

• What goes in — G​limpse takes video and images and a question in natural language. Intern-Decision-4B takes a shared state, a schema of explicitly named questions, and up to eight images.

• What comes out — G​limpse returns prose plus a trace of where in the clip it looked. Intern-Decision-4B returns a distribution over up to 62 symbols, which is a number per option and nothing else.

• Output cost — one forward pass against a token stream. G​limpse bills $3.00 per million output tokens and its mechanism deliberately spends extra tokens looking again, so the trace is the receipt. Intern-Decision-4B produces a fixed handful of logits per request, which makes output cost nearly constant rather than proportional to how hard the question is.

• Size — 4.54 billion parameters, published, downloadable. Unpublished for G​limpse, which states only that it is "substantially smaller and cheaper" than the frontier models in its own benchmark table.

• Weights — open for every member of the Intern-Decision family. None for G​limpse, which exists only behind E​lorian's API and chat interface.

• Independence — the two are equal here, and equally weak. G​limpse's benchmark rows are E​lorian's own, unreproduced, run at its highest thinking effort against rivals evaluated below theirs. Intern-Decision-4B has no published evaluation of any kind that this article could find; the numbers in its configuration are attention geometry, not accuracy.

The limit hiding in the tokeniser

Here is the number that decides whether Intern-Decision-4B is usable for a given job, and it is not on any headline.

The underlying Q​wen 3.5 4B tower declares a maximum position embedding of 262,144 tokens. The decision engine wrapped around it is constructed with max_length=8192, and inputs above that are rejected outright rather than truncated. On its own that is a defensible engineering choice — silent truncation turns a long context into a wrong answer with no error. But the eight images count against the same 8,192 budget. In a workflow that hands the model a screenshot set plus a shared state describing a case, the practical ceiling is reached far sooner than 262,144 suggests, and the failure is a hard rejection rather than a degraded answer.

Anyone planning to use Intern-Decision-4B for document review should test that boundary before designing around it. The demo Space, the collection and the repository were all inaccessible when this was written, so the only way to establish the real ceiling is to construct inputs and watch where the engine refuses them.

G​limpse has the mirror-image problem in the same place. Its context window has never been published, and the model's headline capability is reading video. A 0.9-second interval fits in any window. A two-hour clip does not, and no published figure says where the line is. In both cases the reasoning capability is described and the container it has to fit inside is the part left to guesswork — a coincidence of omissions rather than a coincidence of design.

What each one is genuinely better at

The concrete case for G​limpse is temporal resolution. Fixed-schedule sampling loses whatever falls between two sampled frames, which is why pan-or-cut, object identity across an occlusion and sub-second events are hard questions rather than easy ones. More parameters at eight frames per second do not recover the missing information; spending extra compute on one interval after the whole clip has been seen does. E​lorian's traces show the model doing exactly that, and the visual scratch space makes the decision inspectable after the fact — you can see whether it ever went back to the interval that decided the answer.

The concrete case for Intern-Decision-4B is that it removes the parsing step from a classification pipeline entirely. Consider an underwriting or triage flow that currently calls a large model with a prompt asking it to reply with one of five labels, then parses the reply, then retries when the model answers with a sentence instead of a label, then treats "probably yes" as if it were a yes. Intern-Decision-4B collapses that to one forward pass that returns five probabilities. No parser, no retry loop, no ambiguity about what the model meant. If the threshold on the third label is 0.7, that is a number in your code rather than a heuristic in your prompt.

The case for neither is worth stating too. If what you need is a paragraph explaining what is happening in a scene, Intern-Decision-4B cannot produce it — it has no generation path at all. If what you need is a threshold decision on the semantic content of an image, G​limpse is an expensive way to get a number you could have got more cheaply, and its vendor's own benchmark table is evidence that rewinding the media does not improve inference about what a scene implies: on E​lorian's Humanity's Sixth Sense evaluation, G​limpse scores 46.36 against GPT-6 Astra's 53.6, on a benchmark E​lorian built with Scale AI.

Putting either one behind a production path

The routing question here is not about price, because the two models are not price-substitutes — one is billed per token, the other is a fixed forward pass, and neither is on our catalogue. It is about what sits in front of a model whose failure mode is a hard rejection.

Intern-Decision-4B refuses inputs above its engine limit. A pipeline that has no fallback for that refusal drops requests. OrcaRouter's answer to that shape of problem is automatic failover: one key in front of 200-plus models, with a routing DSL deciding which request goes where, so a request the decision engine rejects can be sent to a generative model that reads the same inputs and writes the answer in prose. The provider's list price is passed through with nothing added per token, so the secondary path costs exactly what its own vendor charges and moves the day that vendor moves its rate. The same layer makes an unproven preview testable without promoting it: put G​limpse on a shadow route, compare its verdicts against a model you already trust, and promote it only when the traces show it going back to the intervals that matter. We serve neither model — E​lorian's own API is where G​limpse lives, and Intern-Decision-4B is a download — so the routing layer is doing work on the shape of the pipeline, not on the availability of these two.

What would change the answer

For G​limpse, three things: an independent harness reproducing a row of E​lorian's table, a published context window, and any disclosure of the base model the company post-trained on. All three are currently blank.

For Intern-Decision-4B, two things. Documentation would be the first — three model uploads in forty seconds with an inaccessible Space, an inaccessible collection and a 404 repository is not a release, it is a snapshot that escaped. A decision engine with no published accuracy figures cannot be calibrated against anything, and a model returning probabilities is worth exactly as much as the trust you can place in them. A raised engine limit or honest guidance on the image budget would be the second, because an 8,192-token ceiling that silently includes up to eight images is a specification surprise waiting to happen in production.

Until those arrive, the split is clean and has nothing to do with which model is larger. If your job is to watch footage and tell a person what happened, with the interval that decided the answer visible in the trace, G​limpse is the only one of the two that can do it, and the free sign-up makes the test cheap. If your job is to classify, score or decide against named options and you need a calibrated number per option rather than a sentence about them, Intern-Decision-4B was built for precisely that and nothing else — accept that it is undocumented, test the 8,192-token boundary early, and do not expect it to explain itself, because it has no mechanism for doing so.

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.