frogmeet

Analysis: Moondream 3 Preview vs 3.1 price and Hub downloads

Same OpenRouter price on this fetch. Different Hugging Face last-month downloads. The gap is not a vision score.

Choosing Moondream 3 Preview or Moondream 3.1 on OpenRouter starts with whether the posted price really matches, then with what Hugging Face downloads do not measure.

On the 2026-10-21 models list both rows take text and image in and return text. OpenRouter describes each as a 9B-total, 2B-active mixture-of-experts vision-language model, and calls 3.1 the successor to Preview. Both descriptions end in an ellipsis in the payload. Same price is this week’s list. Hub downloads are repository counters. This fetch has no vision-benchmark scores, and this note does not invent which model sees better.

Match the priceRead downloadsDo not rank
$0.30per million input tokens, both rows
85383Preview repo downloads
116293.1 repo downloads

Snapshot on 2026-10-21

FieldMoondream 3 PreviewMoondream 3.1
OpenRouter idmoondream/moondream3-previewmoondream/moondream3.1-9b-a2b
Hugging Face idmoondream/moondream3-previewmoondream/moondream3.1-9B-A2B
Context3200032000
Input (USD / 1,000,000 tokens)0.300.30
Output (USD / 1,000,000 tokens)2.502.50
Max completion40964096
OpenRouter created (UTC)2026-10-15 22:13:052026-10-15 22:12:53
Hub createdAt2025-09-112026-06-30
downloads8538311629
likes69289
trendingScore310

What each number can say

One list, one price

Prompt and completion match. Context is 32000 on both. Max completion is 4096. Defaults are temperature 0.2 and top_p 0.9, with reasoning off. That says nothing about latency, and nothing about next week’s price.

More downloads is not “sees better”

The model pages label downloads as Downloads last month. The figures match the API: 85383 and 11629, a gap of 73754, about 7.3 times (85383÷11629). Likes are 692 and 89. trendingScore in the same response is 3 and 10. The directions disagree, so this note picks no winner.

Check it again yourself

  1. 01
    Match the two official ids

    On https://openrouter.ai/api/v1/models find moondream/moondream3-preview and moondream/moondream3.1-9b-a2b. A Hub search for moondream3 also returns third-party copies and quantizations. Leave those downloads out of these two rows.

  2. 02
    Scale per-token USD to per million

    Multiply prompt and completion by 1,000,000. 0.0000003 becomes 0.30. 0.0000025 becomes 2.50. Do not use a different unit.

  3. 03
    Keep the two clocks apart

    Hub createdAt is the repo date: Preview 2025-09-11, 3.1 2026-06-30. OpenRouter created is the listing object: 1792102385 and 1792102373, that is 2026-10-15 22:13:05 UTC and 22:12:53 UTC. The Preview listing is 12 seconds later.

  4. 04
    Stop at fetched facts

    Both model cards name query, caption, point, and detect. The Preview card’s local example uses Transformers and trust_remote_code. The 3.1 card names Photon (pip install moondream, model moondream3.1-9B-A2B) and says local runs target NVIDIA Ampere or newer and Apple Silicon. If a vision score was not in this fetch, leave it out.

# OpenRouter ids, fetched 2026-10-21
moondream/moondream3-preview
moondream/moondream3.1-9b-a2b

Ids, licenses, and the cut-off blurbs

Case of the repo id
The 3.1 OpenRouter id is lowercase moondream/moondream3.1-9b-a2b. hugging_face_id points at moondream/moondream3.1-9B-A2B. Preview is moondream/moondream3-preview on both. canonical_slug values are moondream/moondream3-preview-20261013 and moondream/moondream3.1-9b-a2b-20261013. This note does not treat that suffix as a ship date.
License names differ
The Preview card names Business Source License 1.1 plus an Additional Use Grant. The 3.1 card names the Moondream Model License. Read the repo text before hosting or reselling. This paragraph is not a substitute.
Blurbs are truncated
Both description fields end in an ellipsis, and knowledge_cutoff is null. OpenRouter states visual question answering, captioning, OCR, visual reasoning, 9B total, and 2B active. The cut-off sentence is not filled in here.

To mark the card differences on one shared board, open a browser room at frogmeet. Export the board before you leave. The room does not host the model and does not store API keys.

Questions

Are the two rows the same price now?

On the 2026-10-21 OpenRouter list, yes. Context is 32000. Input is $0.30 and output is $2.50 per million tokens. Pull the list again on the day you call.

Does the higher download count mean Preview is better?

No. 85383 and 11629 match Downloads last month on the model pages. Likes and trendingScore are Hub counters too, and trendingScore points the other way. This page cites no vision benchmark.

Why is OpenRouter created newer than the repos?

It is the listing timestamp, not the weight’s birthday. On the Hub the Preview repo is older (2025-09-11) and the 3.1 repo is 2026-06-30. Using OpenRouter created to call Preview newer swaps those clocks.

Start a room