Glimpse
Glimpse is Elorian's first foundation model — a research-preview "visual thinking" system (released 16 October 2026, with OpenRouter listing on 17 October) that reasons directly over images and video, re-inspecting regions and frames mid-trace rather than converting pixels to a one-shot text description. It ships a 262,144-token context at $0.60/$3.00 per 1M input/output tokens and publishes a compact, vendor-run benchmark card led by CharXiv-R 92.90 and BabyVision 95.36.
Empirical Evaluation Results
Architectural Profile & Capabilities
Glimpse (elorian/glimpse) is the debut model from Elorian, a Palo Alto startup, positioned as a "visual-thinking foundation model." Elorian's announcement is dated 16 October 2026; OpenRouter added the endpoint on 17 October 2026, and it is additionally served via Together AI — making it one of the first third-party visual-reasoning specialists to be broadly routable within days of launch. Capability shape. Unlike caption-and-reason pipelines, Glimpse keeps visual state inside the reasoning loop: it crops and zooms image regions, re-extracts short video intervals at denser frame rates, compares frames across time, and can execute code to measure or diff visual content before committing to an answer. Elorian frames the target capability set as image reasoning, video reasoning and short-form video captioning, open-vocabulary counting, scientific-chart reasoning, manipulation-event understanding, and embodied/physical-world reasoning — i.e. a robotics/engineering/industrial-inspection alignment rather than a document-OCR alignment. Architecture disclosure. Elorian has not published a technical report or detailed model card. Parameter count, active-parameter count, attention mechanism (MLA vs GQA), visual encoder, video tokenization scheme, and whether the backbone is dense or sparse MoE are all undisclosed; third-party summaries describe Glimpse as Elorian post-training an unnamed open-weights base model, which is consistent with the unusually narrow benchmark card and the model's non-release of weights. Any claim of a specific parameter count or MoE routing topology for this variant should be treated as speculation. What is documented is behavioral: a deliberate, multi-pass test-time reasoning regime (Elorian reports all published scores at maximum thinking effort) with tool-assisted visual re-inspection, which implies long output-token traces per query. Commercial positioning. At $0.60/1M input and $3.00/1M output, Glimpse lists at roughly 6% of GPT-6 Astra's price on the same input column. Its own launch table (12 rows) puts it ahead of Astra on 4 rows, level on 1 (ERQA), behind on 6, and holds one row (E.T. Bench) where Astra has no score. Independent summaries stress that the headline CharXiv-R lead over Astra is only ~0.8 points and that Astra still leads Video-MME v2 (64.1 vs 56.68) — cheap, not frontier-dominant, is the accurate read. Benchmark posture for this dossier. Eight scores are explicitly attributed to the Glimpse variant by Elorian's launch materials and reproduced by independent trackers: CharXiv-R 92.90, BabyVision 95.36, TOMATO 93.60, ERQA 83.75, FSC-147 83.36, WGO Bench 42.70, UGC-VideoCap 82.66, and Video-MME v2 (with subtitles) 56.68. These are vendor self-evaluations and are flagged "not independently verified" by the reproducing sources; no MMMU, MathVista, OSWorld, SWE-bench, or text-only frontier scores are published for this variant, and this dossier deliberately reports none. In the October 2026 landscape — GPT-6 Astra, Claude Opus 5 / Sonnet 5, Gemini 3.8 Pro/Flash, DeepSeek V4.x, Grok 4.6, Qwen 3.8, Kimi K3 — Glimpse competes on cost-per-visual-token and on spatial/embodied task families, not on general-purpose agentic or coding leaderboards. Operational note. Because every crop, frame re-extraction, and code-execution step re-enters the context as image/tool tokens, effective cost per answer can exceed the headline $0.60/1M input several-fold on hard visual items — the same tradeoff Elorian itself flags as "looking closely costs extra tokens." No measured TPS, TTFT, or image-encode latency figures are published for this endpoint, so throughput must be established by your own probe before production sizing.
Elaboration is intentionally bounded by what is published. Elorian discloses: (1) a 262,144-token context window, the largest among its listed models; (2) text + image + video input, text output; (3) a deliberate test-time reasoning loop that can crop image regions, re-extract video frames at denser sampling, compare frames over time, and run code to measure visual content — a tool-augmented visual scratch space rather than pure latent-chain reasoning; (4) all reported scores taken at maximum thinking effort. Elorian does NOT disclose total or active parameters, backbone family, attention variant (MLA/GQA/MHA), visual encoder, video tokenizer, pretraining corpus, post-training/RL recipe, or whether the backbone is dense or MoE; several third-party summaries note the weights are not open despite an open-weights lineage claim. isMoe is therefore set to false only because no MoE routing has been evidenced, not because density has been confirmed — treat the field as 'unknown'. The identifiable architectural signature for engineering purposes is architectural-at-inference: extended multi-pass reasoning traces with recurring visual tool calls, which drives token accounting, latency, and context burn more than backbone topology does.
Recommended Workloads & Primary Use Cases
- •Vendor-only evaluation, unaudited. All eight published scores are Elorian self-reports at maximum thinking effort; reproducing trackers explicitly label them unverified and note the CharXiv-R edge over GPT-6 Astra is only ~0.8 points (and Astra leads Video-MME v2 64.1 vs 56.68). Do not treat this card as leaderboard-validated.
- •Research-preview status and undocumented serving envelope. No measured TPS, TTFT, image-encode latency, rate limits, concurrency caps, or deprecation policy are published for elorian/glimpse; routing runs through OpenRouter and Together, so latency and quantization can differ by provider. Pin provider order and run your own p50/p95 probe before committing SLAs.
- •Image tokens dominate cost, and re-inspection multiplies them. The $0.60/1M input figure is per text-equivalent token — high-resolution images, re-crops, dense video frame extraction, and any code-execution round trips all re-enter context, so effective cost per hard visual answer can exceed nominal pricing several-fold. Budget on cost-per-resolved-item, not per-token.
- •Weak on long-horizon video and multi-step manipulation inference. Video-MME v2 (with subtitles) at 56.68 and WGO Bench at 42.70 are the two soft spots on its own card — long-duration video QA, temporal stitching over minutes, and fine-grained action/manipulation reasoning degrade first, while short clips and static charts stay strong; chunk long video and verify before trusting multi-minute temporal answers.
- •Undisclosed architecture blocks capacity planning and troubleshooting. With no parameter count, backbone identity, or context-attention variant published, you cannot predict KV-cache behavior, quantize knowingly, or reason about degradation onset; the 262,144-token ceiling is a nominal figure with no published effective-context or needle-in-haystack validation at long range. Also expect prompt/format sensitivity inherited from the unnamed base model — validate tool-call and image-block formatting empirically rather than assuming OpenAI-style schemas.
Calls to Glimpse are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Elorian.