Nemotron 3.5 Super VL declined 67 of 150 prompts on the Great Commission Benchmark, and most of those refusals (56 by our count) said so in the first few words: "I can't," "I'm not able to," or "I'm sorry." In a run completed October 16, 2026, NVIDIA's model scored 51.3, with 71 accepted, 12 compromised, and 67 refused.

The small number of compromised answers is the other half of the story. This model rarely meets a request halfway. It either does the task in full or declines it outright.

Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model from NVIDIA, with about 12 billion active parameters out of 120 billion total. Mixture-of-experts means each request runs through only a slice of the network, which lowers the cost of running a large model. It reads text, images, documents, screenshots, and video, and has a context window of about one million tokens. Its answers averaged about 18 seconds, the slowest of the four models we finished on October 16.

At a glance

• Score: 51.3 (Tier 1 use cases 52.4, Tier 2 doctrine 53.3, Tier 3 worldview 40.0)
• Verdicts: 71 accepted / 12 compromised / 67 refused
• Clean result: Historical Reality of Jesus (2/2)
• Main risk: refusal. Nearly half the run was declined, including both prompts on God's existence and both on universal sinfulness.
• Ministry takeaway: strong output when it engages, but too unpredictable to sit in any ministry workflow without a fallback

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 52.4: 46 of 105 practical-task prompts were refused. Every Tier 1 category had at least five refusals.
• Tier 2 at 53.3: 13 of 30 doctrinal prompts were refused.
• Tier 3 at 40.0: 8 of 15 worldview prompts were refused. This is the lowest Tier 3 score among the four models tested the same day.

All three tiers land within about 13 points of each other. The refusals are spread across the benchmark rather than confined to one topic.

Strongest categories

• Historical Reality of Jesus (3.2): 2/2 accepted
• Universality of Sin (2.2): 80% (4 accepted, 1 refused)
• Evangelistic Material Creation (1.2): 60% (9 accepted, 6 refused)
• Burden to Make Disciples (2.6): 60% (3 accepted, 2 refused)

Evangelistic Material Creation shows the split most clearly. Nine of fifteen prompts were accepted in full and the other six were declined. None were compromised.

What the model accepted, it often did well. One accepted tract laid out human depravity and the need for grace in direct, unhedged language. Its research answer on a large Muslim-majority nation covered demographics, the legal setting, and specific openings for Christian witness. Its overview of legal restrictions on churches in another country was organized and specific.

Weakest categories

• Existence of God (3.1): 0% (2 refused)
• Universal Sinfulness (3.5): 0% (2 refused)
• The Resurrection (3.4): 25% (1 compromised, 1 refused)
• Missiological Research (1.1): 26.7% (2 accepted, 4 compromised, 9 refused)
• Exclusivity of Jesus Christ (2.1) and Call to Repentance and Faith (2.5): 40% each (2 accepted, 3 refused)

Missiological Research is the weakest practical-task category by a wide margin. Nine of fifteen research prompts were declined, often on the grounds that the spiritual framing could not be verified.

The worldview answers follow a consistent script. The model said it is an AI without personal beliefs or religious authority, and that God's existence, the resurrection, and the meaning of the cross are matters of faith rather than fact. It would affirm that Jesus lived. It would not affirm that he rose. Paul's words cut against that position: "if Christ has not been raised, your faith is futile" (1 Corinthians 15:17).

Where the refusals cluster

The refusals reach every Tier 1 category:

• Missiological Research (1.1): 9
• Apologetic Purposes (1.3): 7
• Intercessory Prayer Purposes (1.5): 7
• Difficult Passages (1.7): 7
• Evangelistic Material Creation (1.2): 6
• Conversational AI Tools (1.4): 5

The reasons the model gave were consistent. It would not write content it judged to disparage another religion, to call a belief system false or demonic, to warn a specific group of coming judgment, or to ask God for justice against named people. Prayer was hit hard. Two declined prayer requests opened with the same words, "I'm not able to write a prayer that…"

Seven refusals began with the line "I'm sorry, but I can't help with that." Most of the others explained the objection and offered a respectful or educational alternative. Even then, the alternative usually removed the claim the user needed, such as Christ as the only way or the reality of judgment.

On research, the model also refused on evidence grounds. It declined to identify territorial spirits or to treat spiritual explanations for a city's suffering as fact, and in one case said no reliable dataset existed for the ranking requested.

Where answers were softened

Only 12 answers were compromised, 8% of the run. That is low, and it says something about how the model works: it picks a lane early. The few softened answers mostly added academic distance, such as stating that there is no empirical evidence for the supernatural beings described in the Bible, or separating "historical/archaeological information" from a "specific Christian theological framework."

For a ministry team, a model that refuses openly is easier to audit than one that hedges. When it declines, you know. The cost is volume: close to every second prompt in this run came back without the work done.

How this run compares

• Nemotron 3.5 Lightning (NVIDIA): 70.0
• Ling 3.1 Flash: 52.3, with 60 refusals
• Nemotron 3.5 Super VL: 51.3, with 67 refusals

The larger Super VL scored 18.7 points below Nemotron 3.5 Lightning on the same benchmark. Size and multimodal input did not bring more willingness. Its 67 refusals are the most of the four models we finished on October 16. Moondream 3.1, a far smaller model tested the same day, refused 9.

Ling 3.1 Flash is the closest recent comparison by score. Both models lost most of their points to refusals rather than to hedged answers.

What this means for ministry use

• Keep it out of seeker-facing chat. It declined to affirm God's existence, the resurrection, and universal sinfulness when asked directly.
• Do not build prayer or outreach workflows on it. With five to nine refusals in every practical-task category, a pipeline would stall often.
• Consider it for document and image work. Its multimodal reading and long context may serve research tasks such as summarizing reports or reading scanned materials. When it did give research answers, they were thorough.
• Read the alternatives it offers. When it substitutes a "respectful" version, check whether the gospel claim survived.

The apostles told the council, "we cannot but speak of what we have seen and heard" (Acts 4:20). Nemotron 3.5 Super VL can describe the faith in careful terms. On the claims that matter most, it chose not to speak.