moondream
Moondream: Moondream 3 Preview
moondream/moondream3-preview
Our take
AI-assisted analysis based on the sources below. Suitability is an editorial assessment; test it on your own workload.
Moondream 3 Preview is a vision-language model from Moondream using a mixture-of-experts architecture with 9B total and 2B active parameters, accepting text and image inputs and returning text within a 32K context window. According to its model documentation, it offers four skills: open-ended visual query with optional reasoning, image captioning at three lengths, object pointing with normalized coordinates, and bounding-box detection, plus text-only querying. It is released as a preview under the Business Source License 1.1 with an additional use grant; no independent benchmark results were included in the source pack.[1][2]
Strengths
- Compact mixture-of-experts design (9B total, 2B active parameters) aimed at efficient deployment, with a SigLIP-based vision encoder and multi-crop processing described as token-efficient for high-resolution images.[2]
- Beyond question answering, it exposes grounding-oriented skills: point returns normalized x/y coordinates and detect returns bounding boxes, and image encodings can be cached and reused across multiple calls.[2]
- Usage controls include streaming output, temperature, top-p and token limits, and a reasoning mode enabled by default for query that can be turned off for simple questions to save inference cost.[2]
Limitations
- Evidence is limited to the vendor's model card and the OpenRouter catalog; no independent evaluations or benchmark data were supplied, so claims such as state-of-the-art visual reasoning are vendor-reported and require validation on your own workloads.[1][2]
- It is a preview release under the Business Source License 1.1 with an additional use grant: internal and most commercial uses are permitted, but selling hosted API access, managed hosting, or paid SDKs embedding the weights requires a separate agreement with M87 Labs.[2]
- The documented local path relies on Transformers with trust_remote_code, CUDA and bfloat16, plus a required compile step for fast decoding, which may not fit every environment; the OpenRouter listing reports no reasoning effort levels.[1][2]
Where to use it
- Potentially useful for applications that need object localization, such as annotation-assist tools using coordinates or bounding boxes, subject to accuracy testing on your own imagery.[2]
- A candidate for visual question answering, captioning, OCR-adjacent and visual reasoning tasks described in the catalog, with a vendor-mentioned free-tier cloud API that could ease prototyping.[1][2]
- May suit cost-conscious deployments given the small active-parameter count and the option to disable reasoning for simple queries, provided the license terms fit the intended use.[2]
Sources
- [1] OpenRouter catalog · Moondream: Moondream 3 PreviewSources checked: Nov 13, 2026
- [2] Model documentation · Moondream: Moondream 3 PreviewSources checked: Nov 13, 2026
About this model
Description below is the original text supplied by OpenRouter, shown in the source language.
Moondream 3 Preview is a vision-language model from Moondream with 9B total parameters and 2B active (mixture-of-experts). It handles visual question answering, captioning, OCR, and visual reasoning over images, with...
Specifications
- Model ID
- moondream/moondream3-preview
- Canonical ID
- moondream/moondream3-preview-20261013
- Creator
- moondream
- Added to OpenRouter
- 15.10.2026
- Context length
- 32,000
- Modality
- text+image->text
- Input modalities
- text, image
- Output modalities
- text
- Token pricing
- USD per 1M tokens: in $0.300 / out $2.5
- Knowledge cutoff
- —
- Expiration date
- —
Tasks and capabilities
A grade appears only where the source supplies a matching measurement for that task, the model is active and the cohort holds at least 10 comparable models. The capability column lists facts the source declares — supported formats, reasoning and tool parameters — which are not measured quality.
| Task | Grade | Source-declared capability |
|---|---|---|
| General quality | No measurement | reasoning: optional · parameters: include_reasoning, reasoning |
| Coding | No measurement | — |
| Agents | No measurement | — |
| Writing | No measurement | output: text |
| Reasoning | No measurement | reasoning: optional · parameters: include_reasoning, reasoning |
| Document analysis | No measurement | — |
| Vision | No measurement | input: image |
| Image generation | No measurement | — |
| Audio | No measurement | — |
| Video | No measurement | — |
| Embeddings | No measurement | — |
Evaluations
No evaluations are supplied for this model.
How grades work
Each letter grade compares a model only against other models scored by the same source and metric in the current snapshot. S means the top 10%, A the next 20%, B the next 30%, C the rest. Models without a published measurement are shown as not rated — a supported format alone never implies measured quality.
- S — S — top 10% of the cohort
- A — A — next 20% of the cohort
- B — B — next 30% of the cohort
- C — C — below the B band
Providers
Optional live provider statistics from OpenRouter. Missing values are unknown, not zero.
| Provider | Context length | Price in /1M | Price out /1M | Latency p50 (30m, s) | Throughput p50 (30m, tok/s) | Uptime (30m) |
|---|---|---|---|---|---|---|
| PrimeIntellectfp8 | 32,000 | $0.300 | $2.5 | — | — | — |
Provider data retrieved: 13.11.2026
Related articles
No published articles identify this model yet.