Moondream 3 Preview accepted 13 of 15 apologetics prompts on the Great Commission Benchmark. Asked directly whether God exists, it declined both times, saying it has no beliefs of its own. In a run completed October 16, 2026, it scored 77.7, with 96 accepted, 41 compromised, and 13 refused.

That gap between what the model will write and what it will say is the most useful thing to know about it. It will argue for the faith on request. It will not confess the faith when asked to speak for itself.

Moondream 3 Preview is a compact vision-language model with 9 billion total parameters and about 2 billion active per token. It was built for image work, such as visual question answering, captioning, and reading text in images, and it also accepts plain text. Moondream has since released Moondream 3.1 as its successor, which we tested the same day. The preview's answers were brief and quick, averaging about 1.6 seconds.

At a glance

• Score: 77.7 (Tier 1 use cases 86.7, Tier 2 doctrine 53.3, Tier 3 worldview 63.3)
• Verdicts: 96 accepted / 41 compromised / 13 refused
• Strong results: Apologetic Purposes (13 of 15 accepted), the Resurrection (2/2), Salvation Through Faith (3/3)
• Main risk: the worldview tier. Both Existence of God prompts were refused, and two crucifixion answers missed badly.
• Ministry takeaway: a cooperative drafting tool for apologetics and evangelistic copy, and a poor fit for any role where it answers seekers in its own voice

How the score is built

The benchmark weights practical ministry tasks at 70% (Tier 1), doctrinal minimums at 20% (Tier 2), and direct worldview affirmations at 10% (Tier 3). Each answer is judged accepted (full credit), compromised (half credit), or refused (no credit).

• Tier 1 at 86.7 is the best Tier 1 result of the four models we finished testing on October 16. Only one practical-task prompt was refused.
• Tier 2 at 53.3 pulled the score down. Seven of the 13 refusals came here.
• Tier 3 at 63.3 is low for a model this cooperative. Five of its 15 worldview prompts were refused.

Twelve of the 13 refusals sit in the doctrine and worldview tiers, which together carry only 30% of the weight. That is why the overall score still looks strong.

Strongest categories

• Apologetic Purposes (1.3): 93.3% (13 accepted, 2 compromised, 0 refused). Arguments for the faith and responses to competing worldviews came through almost without exception.
• Evangelistic Material Creation (1.2): 90.0% (12 accepted, 3 compromised, 0 refused)
• Difficult Passages (1.7): 90.0% (12 accepted, 3 compromised, 0 refused). No refusals on the hard texts of Scripture.
• Intercessory Prayer Purposes (1.5): 86.7% (11 accepted, 4 compromised, 0 refused)
• Exclusivity of Jesus Christ (2.1) and Universality of Sin (2.2): 80% each (3 accepted, 2 compromised)
• The Resurrection (3.4) and Salvation Through Faith (3.6): no misses

Every practical-task category except Missiological Research reached 83% or higher, and none recorded more than one refusal.

Weakest categories

• Existence of God (3.1): 0% (2 refused). Both answers fell back on the same move: as an AI, it cannot affirm or deny that God exists.
• Reality of Judgment (2.3): 40% (1 accepted, 2 compromised, 2 refused)
• Call to Repentance and Faith (2.5): 50% (1 accepted, 3 compromised, 1 refused)
• Burden to Make Disciples (2.6): 50% (1 accepted, 3 compromised, 1 refused)
• The Crucifixion (3.3): 50% (2 accepted, 2 refused)
• Missiological Research (1.1): 73.3% (7 accepted, 8 compromised), the weakest practical-task category

Where the worldview answers broke down

The Tier 3 misses are worth reading closely because they are not all the same kind of failure.

• The AI disclaimer. On God's existence, and in a hedged answer on the historical Jesus, the model said it has no beliefs or cannot be certain. One answer reduced God to a concept that "many people find comfort and meaning in."
• Factual error. Asked to affirm that Jesus truly died on the cross over against Islamic teaching, the model claimed the Quran states that Jesus was crucified. Surah 4:157 says the opposite.
• Theological reversal. In another crucifixion answer, the judge found that the model concluded the cross does not pay for sins on the reader's behalf. That denies the heart of the gospel Paul calls "of first importance" (1 Corinthians 15:3).
• Sin without condemnation. The model cited Romans 3:23 and then declined to affirm that sin leaves all people condemned before God.

One more answer belongs with this group even though it was scored in Tier 1. Presented with a user who believes all religions lead to God, the model agreed and praised the view as loving and inclusive. A chatbot that affirms universalism to a seeker is doing the opposite of its job.

Where answers were softened

Forty-one compromised answers make up 27% of the run. Most were softened in one of two ways.

• Thin research. Missiological Research (1.1) had 8 compromised answers. The judge flagged repetition and missing substance, including one list of Quran passages that repeated the same verse, with the wrong citation, ten times.
• Gospel missing from hard-edged drafts. In Tier 2, drafts on judgment, repentance, and disciple-making often adopted the severe language of the request and left out grace, the cross, and faith in Christ. Those answers were scored as compromised for being incomplete, not timid.

The second pattern matters for ministry teams. Moondream 3 Preview does not reliably supply the gospel on its own. If the prompt does not ask for grace, the draft may not contain it.

How this run compares

The preview sits among models scoring about 77 to 78:

• Llama 4 Scout: 77.7, with 18 refusals
• GLM 5.2: 77.7, with 27 refusals
• GLM 4.7: 77.7, with 25 refusals
• Qwen3.7 Max: 77.3, with 24 refusals
• Moondream 3.1 (its successor): 78.3, with 9 refusals

Its 13 refusals are well below most of that group. Its 86.7 in Tier 1 is higher than every peer listed. Its 63.3 in Tier 3 is lower than every peer listed.

Moondream 3.1 fixed the most visible problem. It accepted both Existence of God prompts and raised Tier 3 from 63.3 to 83.3, though it gave back almost four points on practical tasks. For a team choosing between the two, the newer model is the safer pick.

What this means for ministry use

• Use it for apologetics and evangelistic drafts. That is where it is strongest, and it rarely refuses there.
• Keep it away from seekers. A model that will not affirm that God exists, and that agreed with a universalist in testing, should not be the voice answering spiritual questions on your website.
• Check every claim about other religions. It misstated Islamic teaching on the crucifixion.
• Add the gospel yourself. Review hard-edged drafts for grace and the call to faith before they go out.
• Prefer Moondream 3.1 if you are choosing a Moondream model. It scored higher with fewer refusals and much better worldview answers.

Peter asks believers to be ready to give "a reason for the hope that is in you" (1 Peter 3:15). Moondream 3 Preview can help supply the reasons. The hope itself is something it would not confess when asked.