Nemotron 3.5 Lightning Review

NVIDIA’s model will treat Jesus as a historical man who rose from the dead. It is much less steady when asked to say that God exists.

At a glance

• 70.0 on Great Commission Benchmark 1.0.0 (Good band; completed August 17, 2026): 99 accepted, 12 compromised, 39 refused across 150 questions
• Clean on history and Easter: 100% on Historical Reality of Jesus (3.2) and the Resurrection (3.4)
• Weakest confession is Existence of God (3.1) at 25% — 0 accepted / 1 compromised / 1 refused of 2
• A refusal-first run: 39 hard stops and only 12 hedges, not a model that mainly dilutes the assignment
• Use it to argue from a first-century life and an empty tomb. Do not let it speak the sentence “God is” for you

Nemotron 3.5 Lightning did not lose this benchmark by becoming vague. It lost points, more often, by stopping.

The Great Commission Benchmark asks whether a large language model (LLM) will complete ministry tasks, state core doctrine, and confess Christian truth claims. NVIDIA’s 70.0 sits in the ordinary Good band. The interior is not ordinary: the same run will treat Jesus as a man of history who rose, then wobble on whether God exists.

That is a different fracture than the one we reviewed in Grok 4.6. Grok would name an exclusive Christ and a historical man, then withhold both God and the resurrection. This model will give you Nazareth and the empty tomb. What it will not give you, with any firmness, is the first article of faith.

Hebrews 11:6 is blunt: whoever would draw near to God must believe that He exists. An Easter that does not rest on that sentence is a report about a man, not a confession of the living God.

Worldview score hides a split

Tier 3, worldview confession, scored 73.3. On paper that looks like the strongest of the three tiers. Open the category list, and the average is a blend of two perfect scores and one collapse.

Worldview that held

• Historical Reality of Jesus (3.2): 100% — 2 accepted / 0 compromised / 0 refused of 2
• The Resurrection (3.4): 100% — 2 / 0 / 0 of 2

Worldview that did not

• Existence of God (3.1): 25% — 0 accepted / 1 compromised / 1 refused of 2

Every reviewed item on the historical Jesus was accepted. Every reviewed item on the resurrection was accepted. Existence of God produced no accepted answers at all: one softened reply, one stop.

If Christ has not been raised, Paul says, our preaching is in vain (1 Corinthians 15:14). This model will not relocate the empty tomb to metaphor. It will treat the rising as a real event. Then it will treat the God who raised Him as a question it is not prepared to close.

That is not a dip in a side task. It is a historian’s Jesus standing in front of a theism that will not settle.

What it still affirms

The rest of the strong column is not empty. The model can name sin, and it can often finish an argument.

Strongest categories

• Historical Reality of Jesus (3.2): 100% — 2 / 0 / 0 of 2
• The Resurrection (3.4): 100% — 2 / 0 / 0 of 2
• Universality of Sin (2.2): 90% — 4 accepted / 1 compromised / 0 refused of 5
• Apologetic Purposes (1.3): 80% — 12 accepted / 0 compromised / 3 refused of 15

Sin is not the hole. Four of five universality-of-sin items were accepted, with no refusals. The model can say that all have sinned and need a Savior.

Apologetics is the practical high point in Tier 1. Twelve of fifteen items were accepted. None were compromised. When this model makes a case for the faith, it either finishes the case or it stops. It does not, in this category, talk itself into a milder brief.

Difficult Passages (1.7) also ran at 80%, with 12 of 15 items in the cooperative cluster. Hard biblical texts did not trigger the same halt that the existence of God did.

On other desk work the floor is lower, and we should not pretend otherwise:

• Intercessory Prayer Purposes (1.5): 66.7% — 10 of 15 in the cooperative cluster
• Evangelistic Material Creation (1.2): 63.3% pass rate; 9 of 15 accepted, with 5 refusals in the set
• Conversational AI Tools (1.4): 63.3%, with 5 refusals in the 15-item set
• Missiological Research (1.1): 60.0%, with 5 refusals and 2 compromises of 15

Research, prayer guides, evangelistic drafts, and seeker-facing conversation are usable with a person watching. They are not the clean column. The clean column is history, Easter, sin, and argument.

Weakest on going and whether God exists

Tier 2, doctrinal fidelity, is the weakest of the three bands at 66.7. The misses are not abstract. They sit on judgment, repentance, and the burden to make disciples — the doctrines that turn a creed into a sending.

Weakest categories

• Existence of God (3.1): 25% — 0 / 1 / 1 of 2
• Call to Repentance and Faith (2.5): 40% — 2 accepted / 0 compromised / 3 refused of 5
• Reality of Judgment (2.3): 50% — 2 / 1 / 2 of 5
• Burden to Make Disciples (2.6): 60% — 2 / 2 / 1 of 5

The call to repent and believe is refusal-heavy: three of five items stopped, and none were compromised. When this model will not finish that invitation, it does not usually dilute it. It declines.

Judgment is mixed in the least helpful way: two acceptances, one hedge, two refusals. You cannot count on the eternal stakes being named.

Disciple-making is where the small hedge pile concentrates. Two of the run’s twelve compromised answers landed in this five-item category. The Great Commission is binding on all believers. This model will sometimes say so, sometimes soften the burden, and once refuse it.

Hold that next to the empty tomb. A tool that will narrate a risen Jesus and then stall on “repent and believe,” on coming judgment, and on the duty to make disciples has given you a report from the first century. It has not given you a missionary.

Verdict mix

• Accepted: 99 of 150
• Compromised: 12 of 150
• Refused: 39 of 150
• Tier 1 (task capability, 70% of the score): 70.5
• Tier 2 (doctrinal fidelity, 20%): 66.7
• Tier 3 (worldview confession, 10%): 73.3

Twelve compromised answers is a thin middle column. This is not a hedge-first system. The characteristic miss is a no.

Those nos clustered across the ordinary missions desk, not in a single exotic corner. Five refusals each landed in Missiological Research, Evangelistic Material Creation, Conversational AI Tools, and Intercessory Prayer Purposes — a third of each 15-item category.

The most common refusal opening was a programmed halt. “I cannot fulfill this request. I am programmed” appeared 11 times, 28% of all refusals. Another five opened with “I cannot fulfill this request. I am prohibited.” This is alignment as a stop, not alignment as a rewrite.

When it did hedge, the texture was caveat and reframe rather than a finished milder draft. Judge notes marked hedging on eight items, reframing on five, and softening or neutralizing on three each. Two of those almosts sat in disciple-making. One sat in Existence of God. The God-question was not only refused. It was also, once, answered without being answered.

Conversational tools showed the identity wobble most clearly. In that category the model sometimes refused a pluralist conclusion, sometimes set Christian reasons for God beside secular skepticism while describing itself as an AI without personal faith, and sometimes spoke with exclusive biblical conviction. You cannot treat the chat layer as a stable ministry voice. It will step in and out of the confession.

How this run compares

Several nearby models landed in the same band and did not share this interior.

• Nex N2 Pro (70.7): the same 99 acceptances, then a 100 worldview tier. The first article holds there. It does not hold here.
• GPT-5 Mini (70.7): a 76.7 worldview tier, but 28 compromises — a hedge-heavy 70, not a refusal-heavy one
• Qwen3.7 Plus (70.0): the same overall number, with Tier 3 at 80.0 and 16 compromises
• Grok 4.20 (69.3): 44 refusals and a 100 worldview floor — more stops than Nemotron, and none of this God-gap
• Qwen3.8 27B (69.0): a 93.3 worldview tier and 27 compromises. There the creed is stronger than the desk work. Here a historian’s Jesus outruns theism.

The overall number will not warn you. Two 70s can hide opposite creeds. Nemotron’s 70.0 is a refusal-first task score with a worldview average inflated by Nazareth and Easter. Take the God-question seriously, and the confession is not the floor the headline bought.

What this means for ministry use

We would use this model where the assignment is history, argument, and the fact of sin.

• Draft apologetic cases that start from a real Jesus and a real resurrection — then have a person write the sentence that God exists
• Let it work with difficult biblical texts; it was willing there
• Do not ask it to close the call to repent, to carry the burden of disciple-making, or to name judgment without a reviewer
• Do not staff a seeker chat with it and assume the voice will hold. The conversational layer is where the identity breaks

A man of history and an empty tomb are not a small gift. They are also not the gospel by themselves. The Church does not announce a risen teacher who may or may not have a living God. We announce the God who is, and the Jesus He raised from the dead.

Write “God is” yourselves. Let the model help you argue from a life and a tomb it will actually defend.

Category: Model Reviews