Introducing Glimpse

TL;DR

Elorian launches Glimpse, a visual-thinking MLLM positioned as a new category of model whose chain of thought runs in visual space — zooming into regions, re-watching short intervals at denser frame rates, and tracking features across time inside a “visual scratch space” the UI exposes to the user. Beyond self-driven perception, Glimpse is trained end-to-end to invoke multimodal tools (code, image transforms, video watchers) and to spawn and coordinate subagents that exchange text, visual artifacts, and internal representations. Elorian reports frontier-adjacent multimodal benchmark numbers (matching or exceeding GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 on four of twelve evaluations) at substantially lower API prices ( $0.60/$ 3.00 per 1M input/output tokens, 94% below Astra), post-trained on an open-weights base with “only a few hundred GPUs” by a small team. Available today via API and chat UI, with calibrated 3D scene reconstruction and direct robot action outputs flagged as roadmap.

Key claims

Glimpse is positioned as a visual-thinking model whose reasoning process re-invokes visual representations — zooming, re-watching at higher FPS, tracking features — rather than reasoning purely over text tokens about visual inputs [§“What visual thinking changes”].
Reports frontier-level multimodal benchmark performance at substantially smaller model size and lower API cost than GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, and Gemini 3.8 Flash: $0.60 input /$ 3.00 output per 1M tokens — 94% below Astra, 85% below Opus 5.5, 70% below Sol, 20% below Gemini 3.8 Flash [§“Model results”, pricing table].
Exceeds GPT-6 Astra on four evaluations and matches it on a fifth: WGO Bench manipulation-event understanding (42.70 vs 39.42), FSC-147 open-vocabulary counting (83.36 vs 82.02), UGC-VideoCap short-form video captioning (82.66 vs 80.32), CharXiv-R scientific chart reasoning (92.90 vs 92.10), and ties on ERQA embodied spatial reasoning (83.75) [§“Model results”, Table].
Trained end-to-end to use multimodal tools (writes code, transforms images, watches video) and to spawn/coordinate subagents that exchange text, visual artifacts, and internal representations through a shared problem [§“What visual thinking changes”, intro].
Four visual-thinking modalities demonstrated on worked examples: follow the event through time (track a stable feature across frames to judge speed; re-watch an ambiguous 0.9 s transition at ~14 fps to distinguish pan from cut on 2 birds leaving a door handle), predict what happens next (physical rollout from a photograph of three women lifting a tractor tire), connect an action to its effect (railway switch moves points), and count precisely (count only open pliers while rejecting closed-pliers and scissors lookalikes) [§“What visual thinking changes” blocks 01–04].
Positioned as a robotics foundation model: ablation case using Glimpse as System-2 planner for the Physical Intelligence π0.5 VLA on “put cream cheese in bowl, then put bowl on top of cabinet” — both Glimpse-planner and Astra-planner complete the task, but Glimpse’s planning cost is $0.035 vs Astra's$ 0.249 per episode [§“A foundation model for robotics”].
Developed by a small team with “only a few hundred GPUs” via post-training on an open-weights base; available today via API and interactive chat UI; UI exposes the model’s visual scratch space (region zooms, re-watched intervals, feature tracks) as it reasons [§“Model results”, §“What visual thinking changes”].
Roadmap: multimodal-native pretraining, unified understand-and-generate models, and direct action outputs for robots. Current build is understanding-only; calibrated reconstruction and robot control are flagged as “work ahead” [§“Toward models that understand visual worlds”].

Method

Glimpse is a post-trained MLLM on top of an open-weights base, further trained end-to-end to interleave three capabilities inside a single thinking trajectory: (a) self-driven perception operations over its input — region zooms, denser-FPS re-watches of ambiguous intervals, feature-tracking across time — exposed to the user as a “visual scratch space” during inference; (b) multimodal tool use including code execution, image transforms, and video-segment watchers; and (c) multi-agent coordination where Glimpse spawns subagents that exchange text, visual artifacts, and internal representations.

Model-architectural details, training data, exact substrate (whether the “visual scratch space” is autoregressive-tagged perception ops in the CoT, external tool actions, or trained attention-based re-inspection), and base-model identity are not disclosed in the launch post. Elorian notes training used “only a few hundred GPUs” by a small team, post-training only — framing the model as evidence that visual-thinking capability is cheap once applied on top of a strong open-weights base. Pricing is $0.60 per 1M input tokens and$ 3.00 per 1M output tokens.

Evaluation harness (as reported in the model-card table footnote): all competitor results run at the vendor’s highest thinking effort with their native agent harness (Astra max + Code Interpreter; Sol max + Code Interpreter; Opus 5.5 max + Code Interpreter; Gemini 3.8 Flash high + Code Interpreter for image or Gemini 3.8 Flash high + Video Agent for video). This is the apples-to-apples baseline Introducing Agentic Vision in Gemini 3 Flash flagged for future frontier-vision reporting — Elorian’s table complies with it.

Results

Twelve-benchmark table (Glimpse vs GPT-6 Astra / GPT-6.1 Sol / Opus 5.5 / Gemini 3.8 Flash):

Robotics: WGO Bench end-to-end manipulation events 42.70 (vs 39.42 / 38.91 / 12.34 / 33.99) — SOTA. ERQA embodied spatial reasoning 83.75 (ties Astra; 82.45 / 82.20 / 81.20).

Visual perception: BabyVision fine-grained visual reasoning 94.33 (vs 97.68 / 97.16 / — / 83.25). HSS scene understanding image+video 46.36 (vs 53.6 / 46.6 / 49.81 / 41.6).

Scientific reasoning: CharXiv-R 92.90 (vs 92.10 / 92.60 / 93.40 / 91.80) — exceeds Astra and Sol, below Opus 5.5.

Video captioning: UGC-VideoCap short-form 82.66 (vs 80.32 / 80.41 / 76.35 / 78.33) — SOTA. DREAM-1K dense video captioning 76.04 (vs 79.05 / 77.93 / 72.05 / 70.00).

Temporal reasoning: TOMATO fine-grained temporal 93.60 (vs 96.02 / 89.02 / 72.44 / 83.02). Video-MME v2 w/ subtitles 56.68 (vs 64.06 / 65.27 / — / 59.58). E.T. Bench event understanding 63.55 (vs — / 59.79 / — / 63.45).

Counting: FSC-147 open-vocabulary 83.36 (vs 82.02 / 82.69 / 73.78 / 67.06) — SOTA. CountQA product-image 82.66 (vs 86.42 / 86.42 / 82.20 / 81.09).

Robotics planner cost comparison on a two-step kitchen task driven by Physical Intelligence π0.5: Glimpse $0.035 / episode vs Astra$ 0.249 / episode (7.1× cheaper, both complete the task).

Not reported: ablations isolating which capabilities (visual scratch space vs tool use vs subagents) drive each benchmark gain; model size; whether the pricing is sustainable or launch-promotional (table notes Gemini 3.8 Flash prices as “introductory”); per-benchmark harness details beyond the aggregate footnote.

Why it’s interesting

Glimpse is the first productized frontier-adjacent MLLM whose explicit pitch is substrate (a) + substrate (c) from Thinking with Modalities — autoregressive tagged perception operations inside the CoT, composed with external tool execution — rolled into one trained model and exposed to the user as a visible “visual scratch space.” The combination resembles the direction Introducing Agentic Vision in Gemini 3 Flash described for Gemini 3 Flash (external Python-driven Think-Act-Observe loop) and the research line Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens (continuous visual tokens interleaved into CoT, supervised by heterogeneous vision experts) explored at Qwen2.5-VL-7B scale — but Glimpse is the first to productize both at a frontier-competitive multimodal-benchmark level with public API pricing. The robotics case is also sharper than most VLM launches: a direct System-2-planner-for-π0.5 comparison vs GPT-6 Astra at 7.1× lower cost is the kind of concrete evidence VLA Models has been accumulating for robot-use agent framings (see also ARC: A Reasoning Recipe for Robot Foundation Models on ARC’s reasoning-recipe approach to the same role).

The strongest reservation is epistemic: no model size, base model, training data, or ablations isolating which trained components (visual scratch space, tool use, subagents) carry which benchmark gains, and the “only a few hundred GPUs” framing invites the obvious question of how much of the result is base-model inheritance vs. post-training novelty. The counting demos sit squarely inside the VLM Perception Failures regime — the same tasks (open-pliers vs closed-pliers vs scissors; FSC-147 open-vocabulary) that visual-scaffolding approaches like Visual Structures Helps Visual Reasoning: Addressing the Binding Problem in VLMs showed are massively improved by input-side overlays with no model training at all. A head-to-head between Glimpse’s trained visual-scratch-space and the training-free scaffold-plus-sequential-scan prompt would be the right ablation to isolate what “visual thinking” is actually buying over cheaper alternatives — none of the three Agentic-Vision-style substrates has yet been compared on a shared benchmark Thinking with Modalities §“Open questions”.