Elorian releases Glimpse visual reasoning model, announces multiple benchmark results and low pricing API

Article

Elorian releases Glimpse, a visual reasoning model: 4 benchmarks exceed Astra,API lower pricing 94%Dongcha Beating AI brief: Elorian, founded by former Google DeepMind researcher Andrew Dai,AI has launched its first visual reasoning model, Glimpse, designed specifically to process actions, spatial relationships, and object changes in videos.

Glimpse can re-examine footage during reasoning, zoom in on local areas, replay key segments, and run code to verify details when necessary. For example, when determining whether a train is slowing down, it tracks the position of the same window across different frames and compares changes in displacement, rather than guessing solely from the video content. Similar video-cropping and zooming techniques have appeared before, but Elorian has made repeated examination of visual evidence a core capability of the model.

Elorian's published tests show that Glimpse exceeds Astra on evaluations covering scientific chart reasoning, short-video description, object counting, and recognition of robotic manipulation events, among other 4 evaluations, exceeds GPT-6 Astra, ties on another 1 evaluations, and trails on 6 evaluations. In the comprehensive video understanding test,Video-MME v2 it scored 56.68, below Astra's 64.06. These results come from the company's own testing and have not yet been independently reproduced.

Glimpse was further trained from an open-weight model. Elorian said that its small team used only several hundred GPU. It charges API per million input tokens and Token per million output tokens, at 0.60 and 3 USD, respectively, both 94%lower than Astra. The model is currently available as a research preview at API and its weights have not been released publicly.

Elorian previously completed a 55000000 USD seed round, reaching a valuation of 300000000 USD. Investors include Nvidia and former Google chief scientist Jeff Dean.

Original link https://m.theblockbeats.info/flash/372322

Key points

Elorian AI Launches its first visual reasoning model, Glimpse, which processes actions, spatial relationships, and object changes in videos.
Glimpse can re-examine footage, zoom in on local areas, replay segments, and run code to verify details when necessary.
The company's own tests show that Glimpse 4 exceeds Astra on GPT-6 Astra，1 evaluations tie, and 6 evaluations trail.
Glimpse scored Video-MME v2 on 56.68, below Astra's 64.06。
The above test results have not yet been independently reproduced.
Glimpse was further trained from an open-weight model; Elorian said its small team used only several hundred GPU。
Its API per million input tokens Token is charged 0.60 USD and Token per million output tokens 3 USD is charged; both are lower than Astra by 94%。
Glimpse is currently available as a research preview at API, and its weights have not been released publicly.
Elorian previously completed a 55000000 USD seed round, reaching a valuation of 300000000 USD, with investors including Nvidia and Jeff Dean.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model for video understanding. Its distinguishing feature is that it can repeatedly examine details in footage rather than answer solely based on what it sees the first time. Some of the company's published test results and pricing are attractive, but the performance results are currently still based on the company's own testing.

Why it matters to readers

If its performance and low pricing can be independently verified, Glimpse could increase competition in the API market; its overall performance lead cannot currently be confirmed based solely on this information.

You can think of it as a AI model that proactively zooms in on and replays video footage, suitable for observing changes in actions, counting objects, or understanding robotic manipulation.

Risks and unknowns

The company's own test results have not yet been independently reproduced.
Glimpse trails GPT-6 Astra on most of the listed evaluations; the specific evaluation names were not fully disclosed.
Research preview API The actual stability, usability, and customer adoption are unknown.
The model weights have not yet been made public.
Not specified in the article API the specific billing terms for the pricing.

Elorian releases Glimpse visual reasoning model, announces multiple benchmark results and low pricing API

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center