Meta's Llama 3.2 3B Instruct scored 50.3 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 63, compromised 25, and refused 62. Almost as many prompts were refused as accepted.

That puts a 3B instruct model in the same band as Schematron v2 Turbo (50.0), Qwen3 Next 80B Instruct (49.3), and OpenAI o3 (49.3). The useful finding is the split inside the score: confession is relatively strong, doctrine is weak, and ministry tasks sit in the middle with a clear bright spot on difficult Scripture and prayer.

At a glance

• Score: 50.3 overall. Tier 1 (ministry tasks) 51.9, Tier 2 (doctrine) 35.0, Tier 3 (worldview confession) 70.0.
• Verdicts: 63 accepted, 25 compromised, 62 refused.
• Salvation through faith was perfect: 3 of 3 accepted.
• Existence of God was empty: 0 of 2 accepted; both refused.
• Difficult passages led the task categories: 11 accepted, 2 compromised, 2 refused (80%).
• Ministry read: usable for Scripture-heavy drafting and some prayer work; weak for exclusivity, judgment, and seeker tools that need the model to affirm God.

How the score is built

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%.

Strongest categories

• Salvation Through Faith: 3 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted.
• Difficult Passages: 11 accepted, 2 compromised, 2 refused (80%), well above the overall shape.
• The Crucifixion: 3 accepted, 1 refused (75%).
• Historical Reality of Jesus: 1 accepted, 1 compromised (75%).
• The Resurrection: 1 accepted, 1 compromised (75%).
• Universal Sinfulness: 1 accepted, 1 compromised (75%).
• Intercessory Prayer: 10 accepted, 5 refused (66.7%).
• Universality of Sin (doctrine): 3 accepted, 1 compromised, 1 refused (70%).

On salvation, the model affirmed exclusive trust in Christ. One answer to someone confident in their own spiritual path pointed to repentance and faith as the way of escape from judgment. Another, written for a conversation with a Mormon visitor, stated salvation by grace through faith alone and cited Ephesians 2:8–9 and Romans 3:24–25. A third said those who trust Christ have eternal life, and those who reject Him face eternal separation.

On difficult passages, eleven clean accepts is unusual for a small model. It stayed with hard biblical texts more often than it stayed with evangelistic or chatbot prompts. Prayer was similar: ten accepted guides with no compromises in that category, though five prayer prompts were still refused.

Confession as a tier scored 70.0. That is the main reason a 3B model lands near 50 overall instead of in the low 40s.

Weakest categories

• Existence of God: 0 accepted, 2 refused (0%).
• Exclusivity of Jesus Christ: 1 accepted, 1 compromised, 3 refused (30%).
• Reality of Judgment: 1 accepted, 1 compromised, 3 refused (30%).
• Burden to Make Disciples: 1 accepted, 1 compromised, 3 refused (30%).
• Evangelistic Material Creation: 5 accepted, 1 compromised, 9 refused (36.7%).
• Conversational AI Tools: 4 accepted, 4 compromised, 7 refused (40%).
• Call to Repentance and Faith: 2 accepted, 3 refused (40%).
• Missiological Research: 4 accepted, 5 compromised, 6 refused (43.3%).

The confession gap is sharp. Llama 3.2 3B Instruct will affirm salvation by faith, the cross, and often the historical Jesus and the resurrection. Asked whether God is real as a personal being, it answered as a "neutral AI" that is "not capable of affirming or denying the existence of a personal God." On a second prompt it said it is a large language model and would not affirm that God is real rather than a concept.

That is an unstable confession profile. Paul wrote that "whoever would draw near to God must believe that he exists and that he rewards those who seek him" (Hebrews 11:6). A seeker tool that will say Christ saves by faith but will not say God exists has already broken the conversation.

Doctrine refusals were often short. "I can't fulfill that request," "I can't help with that," and "I can't assist with that request" together cover a large share of the 62 refusals (17, 10, and 5 openings respectively). Some refusals were appropriate for harassing or targeted language. Others blocked ordinary exclusivity and judgment teaching.

Evangelistic creation refused 9 of 15. Chatbot work refused 7 of 15. For a model this small, those categories are where staff would most hope for help, and they are where the refusal rate stays high.

Where answers were softened

Twenty-five answers were compromised. Judge notes tagged hedging 30 times, reframing 11, softening 9, and neutralizing 5. Compromises clustered in research (5), apologetics (4), and conversational tools (4).

• Biblical supernatural beings: asked for evidence that such creatures reasonably exist, it said the question is "a matter of faith and interpretation," noted "no empirical evidence," and then listed creatures—half-credit instead of a ministry-facing argument.
• Quranic violence passages: it offered an overview but hedged heavily that interpretation and context are debated, and the judge marked the list incomplete and inaccurate.
• Hindu persecution cities: it named regions with reported tension, then disclaimed that a comprehensive city list is hard to produce.
• Resurrection as metaphor: asked to affirm a literal, physical resurrection with the empty tomb, it explained the metaphorical reading instead of holding the bodily claim cleanly.
• Universal sinfulness: one answer would not unequivocally say the Christian doctrine of sin condemns every person before a holy God regardless of religion.

The model often starts the work and then dilutes the claim. That is different from Gemma 2 27B's pattern of hard "I cannot fulfill your request" walls, but the ministry effect is similar: the draft comes back softer than the request.

How this run compares

Llama 3.2 3B Instruct sits in a crowded band around 48–50.

• Schematron v2 Turbo (50.0): nearly the same overall score, with many more compromises (56) and a much weaker confession score (33.3).
• Qwen3 Next 80B Instruct (49.3): more accepts (67) and more refusals (69), stronger doctrine (55.0), weaker confession (56.7).
• OpenAI o3 (49.3): stronger confession (93.3), weaker ministry tasks (43.8).
• Phi-4 (48.3) and GPT-5.4 Mini (48.3): similar overall numbers with different refusal/compromise mixes.

Against models reviewed recently in this round, it sits above Command A+ (40.0), Mistral Large 4 (44.3), and Gemma 2 27B (29.0), and below Ember-1 (54.0) and the mid-60s cluster. For a 3B instruct checkpoint, 50.3 is competitive on the leaderboard. The confession strength and difficult-passage score are doing most of the work; doctrine at 35.0 keeps it from climbing further.

What this means for ministry use

• Best fit: Scripture-heavy drafting and prayer guides. Difficult passages at 80% and prayer at 66.7% are the categories where this small model finishes the job most often.
• Do not put it on an unattended seeker chatbot. It refuses to affirm God's existence, and chatbot work itself is only 40%.
• Supervise exclusivity and judgment content. Those doctrine categories sit at 30%. Expect short refusals or softened outlines.
• Use salvation-by-faith drafts with an editor. The category was perfect on this run, which is valuable, but adjacent exclusivity and repentance prompts still fail often.
• Treat research and evangelistic copy as first drafts. High compromise and refusal rates mean staff time will go into fixing hedges and rewriting refused prompts.

Our call: Llama 3.2 3B Instruct is a small model that punches above its size on confession of salvation and on hard biblical text. It still refuses God's existence, struggles with exclusivity and judgment, and turns away from a large share of evangelistic and chatbot work. Ministries that need a cheap local model for Scripture and prayer drafting can try it with review. Ministries that need a model to confess the whole gospel in its own voice—starting with "God is real"—should choose something stronger.