Moondream 3 Preview lists on OpenRouter: 9B MoE vision model, 2B active, 32K context, $0.30/$2.50

Moondream 3 Preview, a vision-language model with 9B total parameters and 2B active per token, is listed on OpenRouter at $0.30 per 1M input tokens and $2.50 per 1M output tokens. It has a 32K-token context window and an optional reasoning mode.

Specifications

Spec	Detail
Developer	Moondream
Architecture	Mixture-of-experts (MoE)
Total parameters	9B
Active parameters	2B
Context window	32K tokens
Input price	$0.30 per 1M tokens
Output price	$2.50 per 1M tokens
Release date (per OpenRouter listing)	Oct 15, 2026
Reasoning mode	Optional

According to the listing, the model handles visual question answering, image captioning, OCR and visual reasoning over images. The page links to model weights. It does not state a license in the text available, so licensing terms are not confirmed here.

Serving performance

Prime Intellect is the only provider shown on the OpenRouter page. It lists the following figures at the same $0.30 / $2.50 pricing:

• P50 latency: 0.33 seconds
• Throughput: 208 tokens per second

OpenRouter's uptime figures for the model were not populated, and the page reports not enough data on apps using it yet. These are OpenRouter's measurements of its listed provider, not independent benchmarks.

What is not disclosed

The listing gives no benchmark scores, so there are no published numbers for VQA, OCR or captioning accuracy to compare against other vision-language models. The training data cutoff, the expert count and the routing configuration are also not disclosed. The "Preview" label suggests the weights and behavior may change before a stable release.

The same page lists Moondream 3.1 as a successor to Moondream 3 Preview. Its listed specifications match the preview: 9B total parameters, 2B active, 32K context, optional reasoning mode and $0.30 / $2.50 pricing. The listing does not say what changed between the two versions.

What this means

The notable design choice is the active-parameter count. A 9B MoE that activates 2B parameters per token should have per-token compute closer to a small dense model, while keeping more total capacity. Whether that translates into accuracy is unknown without benchmarks, and Moondream has not published any in this listing.

The pricing is asymmetric. Output tokens cost about 8.3 times as much as input tokens. Image-heavy workloads such as OCR and captioning, which send many input tokens and return few, are priced more favorably than workloads that produce long outputs. Enabling the reasoning mode will likely raise output token counts and therefore cost, though the listing does not quantify this.

With a single provider listed, there is no routing redundancy yet. Teams considering production use should wait for multiple providers or run their own deployment from the released weights. The appearance of Moondream 3.1 under the same listing also suggests that the preview may be short-lived, so developers should plan for version changes.