Glimpse vs MiniCPM5 2B: Seven Weeks of Paper Trail and Nine-Tenths of a Second

MiniCPM5-2B and E​lorian G​limpse are both small models claimed to beat things much larger than themselves, and they make the claim in opposite currencies. MiniCPM5-2B, shown by M​odelBest and O​penBMB at WAIC on 19 July 2026 and finally given weights on 6 September 2026, comes with a card averaging 53.9 across its own test set against larger open models topping out at 51.1. G​limpse, released 16 October 2026, comes with a 0.9-second window — a first pass at roughly 8 frames per second, then a re-scan of nine-tenths of a second at around 14 frames per second — and a set of benchmark rows that are the vendor's own, unreproduced, and notably mixed. One of these has seven weeks of public paper trail behind it and a downloadable artefact. The other has a mechanism and almost no numbers.

The comparison is not really two-versions-of-the-same-thing, because MiniCPM5-2B is a general reasoning model that also reads images while G​limpse is a hosted visual-reasoning preview that does almost nothing else. The reason to read them together is that both are arguments about what small models can do — and the two arguments fail in different, instructive ways.

The wait, and what the wait produced

MiniCPM5-2B's timeline is worth reconstructing, because it is the part of the story a release-date search would miss.

The model was unveiled on 19 July 2026 at the World Artificial Intelligence Conference, where the headline capability was a 512,000-token context window. The weights did not appear until 6 September, with a changelog following on 7 September. Between those two dates the model existed as a demonstration and a set of numbers.

What shipped is not what was advertised, and the gap is checkable. The released configuration sets max_position_embeddings to 131,072 with no rope_scaling block, which is 131,072 tokens — a quarter of the WAIC figure. That is not dishonesty so much as the ordinary difference between what a research team shows on stage and what an engineering team is willing to sign off in a config file, but it is the number that matters, because it is the one your deployment inherits.

What did ship is unusually complete for a model in this size class. Alongside the final BF16 checkpoint — RL plus OPD merged — there are SFT, midtrain and base stages published separately, GGUF, MLX and GPTQ builds for local runtimes, a DSpark draft model for speculative decoding, and the training corpora themselves released as the UltraData family. Day-zero support on nine chipsets came through FlagOS, and the card makes a point of stating that the implementation uses no custom kernels and no model-code fork. That last detail is the one that makes the rest credible: a model that runs on stock tooling is a model that runs where you are, not just where the authors' cluster is.

G​limpse's timeline is compressed by comparison and its paper trail is thinner. Released 16 October 2026, one day before this article, as a research preview reachable through E​lorian's API and chat interface with a free sign-up. No weights, no publication of the base model it was post-trained on, no artefact of any kind. The company came out of stealth in April 2026 behind Andrew Dai, a former G​oogle researcher, with $55 million from Striker Ventures, Menlo Ventures and Altimeter and participation from N​VIDIA, 49 Palms and Jeff Dean. The model is the first thing it has shipped.

The numbers, and what each set covers

• Parameters — MiniCPM5-2B is 2,516,756,480 total, of which 1,981,982,720 are non-embedding; 42 layers, group-query attention across 16 query and 2 key-value heads, served through a stock LlamaForCausalLM class. G​limpse's parameter count is unpublished, and its launch post says only that it is "substantially smaller and cheaper" than the frontier models it benchmarks against.

• Artefact — Apache 2.0 weights plus SFT, midtrain, base, GGUF, MLX and GPTQ variants for MiniCPM5-2B. None for G​limpse, which offers a hosted endpoint and nothing to download.

• Context — 131,072 tokens as configured for MiniCPM5-2B, against a 512,000-token figure touted at the WAIC demonstration. Unpublished for G​limpse — the more consequential blank of the two, since its defining capability is reading video.

• Modality — images in, text out for MiniCPM5-2B. Video and images for G​limpse, with the ability to return to an interval after a first pass.

• Evaluation — MiniCPM5-2B's card reports LiveCodeBench v6 at 69.1, AIME 2025 and 2026 both at 86.5, MATH-500 at 94.6, MMLU-Pro at 70.8, LongBench v2 at 43.7, BFCL v4 at 66.6 and SWE-bench Verified at 46.4, with some rows daggered as carried over from an Artificial Analysis release rather than measured in-house. G​limpse's rows are E​lorian's own, against GPT-6 Astra, with a footnote disclosing that several competitors were run below maximum thinking effort while G​limpse ran at its highest.

• Price — $0.60 per million input tokens and $3.00 per million output for G​limpse, with the rewind billed as output tokens, so the trace is the receipt. No rate card for MiniCPM5-2B, because the deliverable is a file.

The claim both of them make, and the different ways it holds

MiniCPM5-2B's central claim is arithmetic and checkable. A 2.5-billion-parameter model, card average 53.9, against larger open models whose best average is 51.1 on the same set. That is a claim of the form "this small thing beats that big thing on these tasks," and it is falsifiable by anyone with a GPU and an afternoon, because the weights are a download and the tests are published.

The mechanism for it is documented in three stages: supervised fine-tuning on roughly 400 billion tokens of deep-thinking data, reinforcement learning with teacher models, and then on-policy distillation that merges sixteen RL experts — five of them agentic — back into one checkpoint. The internal numbers attached to that merge are a 10.96-point gain on reasoning and general tasks and 6.96 on agentic ones relative to the pre-merge state. Those are the vendor's own measurements of its own pipeline, and they are the kind of claim a third party can now go and check, which is the point of shipping the intermediate checkpoints as well as the final one.

G​limpse's central claim is not arithmetic. It is a behaviour: that returning to a sub-second window after a first pass produces better answers on temporal questions than sampling on a fixed grid ever could. The supporting evidence is a set of worked traces and a benchmark table that splits almost evenly — ahead 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); 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).

Both claims are plausible. Only one of them can currently be checked by someone outside the company that made it.

What the two disclosures have in common

The MiniCPM5-2B card carries daggers on a subset of its rows, marking them as figures taken from an Artificial Analysis release rather than measured by the authors. Read that as a disclosure rather than a defect. The alternative — quietly folding someone else's measurements into your own table without a marker — is common and worse.

Both models are, on the evidence available, unverified in the same specific sense. G​limpse's table is vendor-run at the vendor's best configuration. MiniCPM5-2B has no independent reproduction of its card average either, whatever the provenance of individual rows. On this dimension the two releases are identical, and it is worth noticing that before either is treated as the reliable half of the comparison.

The stronger difference is what happens next. Anyone can download MiniCPM5-2B and run their own evaluation this afternoon, on their own data, without asking anyone. Nobody can do that with G​limpse. The trial is free and the traces are inspectable, but the evaluation is bounded by what E​lorian chose to show and the endpoint is the only place the model exists.

Where custody decides the question

This is the dimension where the two models are not close, and it has nothing to do with capability.

MiniCPM5-2B's deliverable is a set of files under Apache 2.0, including a GGUF build and an MLX build. That means the model runs on a laptop, the images never leave the machine, and the deployment has no vendor in its request path — no availability window, no retention question, no rate limit, no price that can move under you. For regulated work with imagery, or for anything where the media cannot cross a network boundary, that is the entire decision, and G​limpse is not a candidate however good the trace looks.

G​limpse's capability exists only behind E​lorian's endpoint. Its distinguishing feature is the visual scratch space, and there is no artefact to relocate. That is not a flaw in the product; it is a property of a research preview whose value is arguably in the mechanism rather than the weights.

It is also where a routing layer is useful and where it stops being useful. OrcaRouter gives you one key in front of 200-plus models with provider list prices passed through and nothing added per token, so a local MiniCPM5-2B and a hosted model can sit on one code path rather than two, with a routing DSL choosing per request and automatic failover keeping a request alive when a lane fails. The hosted half of that mix, if the vendor cuts its price, is live on our side the same day rather than at the next renewal. What routing cannot do is move custody: if an endpoint is not on your hardware, no routing rule puts it there. Neither MiniCPM5-2B nor G​limpse is on our catalogue — one is a download and the other is E​lorian's own API and chat interface — so this describes how we route, not a claim to serve either model.

The decision, with the accounting attached

MiniCPM5-2B and G​limpse are not competitors, and treating them as such wastes the comparison. One is a 2.5-billion-parameter general model you can own. The other is a hosted preview that does one unusual thing.

There is no shared scoreboard between these two, and no way to build one. MiniCPM5-2B's card measures coding, mathematics, instruction following and long-context retrieval. G​limpse's table measures video and image perception against GPT-6 Astra. They share no task, so the only comparisons that matter are against your own workload. Three questions settle it faster than a table would.

Does the model have to run on your own hardware, or does the licence have to be permissive? Then MiniCPM5-2B is the only candidate here. Download the GGUF or MLX build, budget for the 131,072-token context the configuration actually sets rather than the 512,000 shown at WAIC, and treat the card's daggered rows as pointers to the original evaluation rather than as measurements by the authors.

Is the failure mode temporal — a camera pan mistaken for a cut, an object that changes identity across an occlusion, an event inside a sub-second window? Then G​limpse is the only one of the two whose stated mechanism addresses it. The trial is free, and the trace is the test: run your own clips and check whether the model ever went back to the interval that decided the answer.

Is the work long-context but not long-video — a document set, a filing archive, an audit trail that has to sit inside one window? Neither model's published evaluation settles that. LongBench v2 at 43.7 is a single number for a whole category of behaviour, and it will not tell you how MiniCPM5-2B performs on your documents at 120,000 tokens. That one has to be measured before either decision is safe.

What would change this page: any reproduction of G​limpse's benchmark rows by someone other than E​lorian, the publication of its parameter count or context window, or a MiniCPM5 checkpoint that closes the gap between the 131,072-token config and the 512,000-token demonstration. None is on the calendar. Until they arrive, one of these models can be tested by anyone tonight and the other cannot — and that asymmetry, more than any benchmark row, is the thing to plan around.

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.