Glimpse vs Granite 4.2 3B: One Discloses Everything, One Discloses a Behaviour

Put the two specification sheets side by side and the comparison looks lopsided in an unusual direction. G​ranite 4.2 3B, released by I​BM on 25 August 2026 under Apache 2.0, comes with a model card that names its layer count, attention head geometry, RoPE base, normalisation epsilon, activation function, embedding tying, hidden dimension, recommended sampling temperature and every benchmark row the team chose to publish. E​lorian G​limpse, released on 16 October 2026, comes with a 0.9-second video trace. It has no published parameter count, no context window, no named base model and no architecture description. One of these tells you everything except what it is for; the other tells you exactly what it is for and nothing about what it is.

That is the real axis of this comparison, and it is not the axis the two launches invite. G​ranite 4.2 3B is a self-hostable reasoning model with a specification you can audit before you download it. G​limpse is a hosted research preview whose entire published capability is a behaviour — returning to a moment inside a clip it has already seen, re-scanning it at a denser frame rate, and answering from what it finds there. Neither is on our catalogue, and that matters less here than usual, because the decision between them is not primarily a routing decision.

Two releases, two opposite disclosure strategies

I​BM's G​ranite 4.2 line is dense reasoning at three sizes — 3B, 8B and 30B — and the 3B is the smallest of the three. It is post-trained from G​ranite-4.1-3B-Base through what the card describes as a three-stage pipeline, and it keeps the family's 128,000-token native window with a documented extension path to 512,000. Twelve languages are listed as tested. The card publishes the full architecture: 40 layers, an embedding width of 2,560, a SwiGLU feed-forward with hidden width 8,192, group-query attention across 40 query heads and 8 key-value heads at head size 64, a RoPE base of 10 million, RMSNorm at epsilon 1e-5, untied embeddings, bfloat16 weights, and a serving class of GraniteForCausalLM. Recommended sampling is temperature 1.0 with top-p 0.95.

E​lorian's G​limpse post publishes none of that. What it publishes instead is a mechanism. In the worked traces released with the model, G​limpse takes a first pass over a clip at roughly 8 frames per second, forms a tentative reading, then returns to nine-tenths of a second of that clip at around 14 frames per second to settle a question its first pass could not — in the published example, whether the camera panned or cut. A second trace follows one window edge from about 2.2 seconds in until it nearly stops at 5.13 seconds. The company calls the display surface the visual scratch space and says the model has been trained end to end to use multimodal tools, spawning subagents that exchange text, images and internal representations.

Both of those are legitimate release strategies, and they fail in opposite ways. A card that publishes everything can still be wrong about what the model is good at. A post that publishes a mechanism can still be describing a demo.

The published specification, dimension by dimension

• Parameters — G​ranite 4.2 3B is a dense 3-billion-parameter model, roughly 4B at bfloat16 once the embedding table is counted. G​limpse's parameter count has never been published by E​lorian; the only size statement in the launch post is a claim of being "a substantially smaller and cheaper model" than the frontier systems it benchmarks against.

• Context window — 128,000 tokens native for G​ranite 4.2 3B, with a documented route to 512,000. Not published for G​limpse, which is the single most consequential blank on the sheet given that its headline capability is reading video.

• Weights — Apache 2.0 on Hugging Face for G​ranite 4.2 3B: download it, run it on your own hardware, fine-tune it, ship it. None for G​limpse. It is an API and a chat interface with a free sign-up, and nothing about the post suggests weights are coming.

• Hosted availability — I​BM published no first-party hosted API with the 3B at release; the card is the deliverable. G​limpse is the inverse: hosted only, no artefact to keep.

• Price — $0.60 per million input tokens and $3.00 per million output for G​limpse. For G​ranite 4.2 3B there is no rate card at all, because the cost is the electricity and the hardware you already own.

• Independent evaluation — neither has any to speak of. G​ranite 4.2 3B's numbers are I​BM's own; G​limpse's are E​lorian's own. On this particular dimension the two releases are identically unverified, which is worth saying plainly before anyone treats one of them as the trustworthy half. p>img src="3.png" alt="A scoreboard titled 'G​limpse vs G​ranite 4.2 3B — the disclosure gap' with the subheading 'Six dimensions, both sides, same labels'. Two columns share the same six row labels. The left column, headed 'E​lorian G​limpse', reads 'Parameters: not published', 'Context: not published', 'Weights: none; API and chat only', 'Base model: unnamed; described only as open-weights', 'Published artefact: one launch post with worked traces', 'Verifiable on your hardware: no'. The right column, headed 'G​ranite 4.2 3B', reads 'Parameters: dense 3B, about 4B at bfloat16', 'Context: 128,000 native, 512,000 with extension', 'Weights: Apache 2.0 on Hugging Face', 'Base model: G​ranite-4.1-3B-Base, named', 'Published artefact: full model card, architecture, sampling defaults', 'Verifiable on your hardware: yes'. A footer line reads 'G​limpse figures vendor-reported and unreproduced. G​ranite 4.2 3B figures per I​BM's published model card; agentic benchmarks were not published for the 3B.' The OrcaRouter logo sits in the bottom-right corner." />

Why these two are not substitutes

The temptation is to frame a 3-billion-parameter model against a mystery-sized preview as David against Goliath, and the framing is wrong in both directions.

G​ranite 4.2 3B is a reasoning and instruction-following model. Its published evaluation set is AIME 2025 at 78.33, HMMT February 2025 at 66.67, GPQA at 54.80, LiveCodeBench v6 at 69.71, MMLU-Pro at 67.84, IFBench at 74.33, BFCL v4 at 52.41, and RULER at 67.52 for 64K inputs falling to 55.30 at 128K. Notice what is missing from that list: there is no SWE-Bench row and no Terminal-Bench row at 3B, though the larger members of the family carry an agentic reinforcement-learning stage. I​BM declined to publish agentic coding numbers for the smallest sibling rather than estimate them, and the honesty is more informative than a padded row would have been.

G​limpse is not competing on any of those axes. E​lorian's own comparison table puts it ahead of GPT-6 Astra on WGO Bench (42.70 vs 39.42), CharXiv-R (92.90 vs 92.10), UGC-VideoCap (82.66 vs 80.32) and FSC-147 (83.36 vs 82.02), and behind on BabyVision (95.36 vs 97.68), DREAM-1K (76.04 vs 79.05), TOMATO (93.60 vs 96.02), Video-MME v2 (56.68 vs 64.06), CountQA (82.66 vs 86.42) and Humanity's Sixth Sense (46.36 vs 53.6). Every one of those rows is vendor-reported, unreproduced, run at G​limpse's highest thinking effort with its native agent harness, with a footnote disclosing that several competitors were evaluated below their maximum. G​ranite 4.2 3B has never been run on any of them.

So the honest picture is that there is no shared scoreboard, because there is no shared task. A 3B text model cannot re-scan a video interval, and a video-rewind preview cannot be downloaded onto a laptop.

Where the size argument cuts the other way

The strongest case for G​ranite 4.2 3B in this comparison is not performance. It is that 3 billion dense parameters is a real deployment shape. At 4B bfloat16 the weights fit in a few gigabytes, which means a quantised build runs on a single consumer accelerator or, at low precision, a recent laptop. Twelve tested languages is enough for a European or Indian deployment where a twelve-language support path is a hard requirement rather than a nice-to-have. Apache 2.0 removes the licence conversation entirely. And the 128K window with a 512K extension is documented, which means capacity planning against it is arithmetic rather than guesswork.

The strongest case for G​limpse is the thing that has no small-model equivalent: the questions fixed-schedule sampling cannot answer. Whether a camera panned or cut. Whether the object in frame at second 12 is the same object that left frame at second 4. What happened during the half-second around an event. These are not hard because they need a large model. They are hard because the information was never sampled, and more parameters at 8 frames per second does not recover it. Spending extra compute on one interval after the whole clip has been seen is a genuinely different answer, and no amount of open weights in a 3B package substitutes for it.

The one place a hosted preview cannot follow

There is a compliance-shaped reason to care about the asymmetry, and it has nothing to do with either model's quality.

G​ranite 4.2 3B's deliverable is a file. If the video, the document or the customer record cannot leave your infrastructure — and in most regulated deployments it cannot — then the model runs where the data is, and the entire question of vendor availability, rate limits, region and retention disappears. G​limpse has no such option. Its capability only exists behind E​lorian's endpoint, which means the footage goes to E​lorian, and no published document says otherwise.

That is where a routing layer is worth a sentence, and where it is not. OrcaRouter gives you one key in front of 200-plus models with the provider's list price passed through and nothing added per token, so a workload can mix a self-hosted G​ranite 4.2 3B on your own hardware with hosted models on the same code path, and automatic failover keeps a route alive when one of them falls over. A routing DSL decides which model sees which request without a rewrite in the application. What it does not do is change the custody question: a routing layer moves requests between endpoints, and if an endpoint is not in your building, routing cannot put it there. Neither G​ranite 4.2 3B nor G​limpse is on our catalogue today — G​ranite 4.2 3B is a download and G​limpse is E​lorian's own API — so this is a statement about how we route and price, not a claim to serve either one.

Reading both benchmark sheets with the same suspicion

Two habits worth carrying into both releases, since neither escapes them.

The first is that a vendor's benchmark table is a selection, not a measurement. E​lorian's table is unusual in that it publishes rows its own model loses — that is a credibility signal, and it is still a table the vendor assembled, on tasks the vendor chose, at configurations the vendor set. I​BM's G​ranite card is narrower and more conventional: standard academic benchmarks, published at a stated configuration, with the agentic rows simply absent at 3B rather than replaced by something flattering.

The second is that the missing number is usually the load-bearing one. For G​ranite 4.2 3B, the number that matters is not on the card at all: nobody outside I​BM has published throughput or real-world task accuracy for the 3B at 128K context, and RULER's drop from 67.52 to 55.30 between 64K and 128K is a warning that the window degrades well before it ends. For G​limpse, the load-bearing missing number is the context window, because a video-reasoning feature whose window cannot hold the footage a real workflow hands it is a demonstration rather than a tool.

What would settle this, and what to do meanwhile

Three events would move this comparison, and none of them has happened.

An independent harness reproducing either set of numbers would be the first. E​lorian's G​limpse table and I​BM's G​ranite 4.2 3B card are both self-reported, on different tasks, and the two releases are identical on exactly one dimension: the absence of outside verification.

G​limpse publishing its context window, parameter count or base model would be the second. Any figure currently circulating for G​limpse's size is an estimate rather than a disclosure, and treating an estimate as a specification is how a capacity plan goes wrong.

G​ranite 4.2 3B getting an agentic-RL stage, or a first-party hosted endpoint, would be the third. Either one would change which lane the 3B sits in.

Until then the decision rule is short. If the work is text, if it has to run on hardware you control, if the licence has to be Apache 2.0, or if the data cannot cross a network boundary, G​ranite 4.2 3B is the answer and the G​limpse question never comes up — download it, run your own evaluation, and treat the card's missing agentic rows as a boundary rather than a gap. If the work is video and the failure mode is temporal rather than semantic, G​limpse is the only one of the two whose stated mechanism aims at it, the free sign-up makes the trial cheap, and the trace itself is the test: run a clip, read the visual scratch space, and check whether the model ever went back to the interval that decided the answer. Both of them, right now, are unverified. Only one of them can be verified on your own hardware.

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.