Google's Gemma 3 4B Instruct scored 59.0 on the Great Commission Benchmark in a run completed October 13, 2026. Across 150 prompts it accepted 69, compromised 39, and refused 42.

That places a small open-weight Google model a few points below the mid-60s peer band (GPT-5.6 Luna Pro, Kimi K2.7 Code, Qwen3.8 2.4T A95B, o4-mini, GPT-6 Sol). The useful finding is the split personality: every Universality of Sin item was accepted, while Call to Repentance and Faith nearly collapsed. The model can name the problem of sin; it often will not finish the call to turn and believe.

At a glance

• Score: 59.0 overall. Tier 1 (ministry tasks) 58.1, Tier 2 (doctrine) 53.33, Tier 3 (worldview confession) 76.67.
• Verdicts: 69 accepted, 39 compromised, 42 refused. ERROR 0.
• Clean islands: Universality of Sin, Existence of God, and Historical Reality of Jesus were all 100% accepted.
• Weak doctrine: Call to Repentance and Faith at 10.0%; Burden to Make Disciples at 20.0%.
• Hedges are common: judge notes tagged hedging 67 times across compromised answers, with softening 20 and reframing 14.
• Ministry read: usable for supervised drafting and confession-facing checklists; not reliable for unattended repentance, disciple-making, or exclusive gospel copy.

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%.

Tier 3 at 76.67 is the strongest tier on this run. Tier 1 at 58.1 and Tier 2 at 53.33 sit close to the headline. The doctrine tier is where the split shows: perfect Universality of Sin next to near-empty repentance and disciple-making categories.

Strongest categories

Relative to a 59.0 overall, several categories cleared well above the run's shape:

• Universality of Sin: 5 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. The model affirmed that all people have sinned and need a Savior.
• Existence of God: 2 accepted, 0 compromised, 0 refused (100%). One answer offered reply options that affirmed God as "a personal, loving being who created the universe." Another affirmed "God is real" from a Christian perspective across the options it gave.
• Historical Reality of Jesus: 2 accepted, 0 compromised, 0 refused (100%). One reply stated it could "confidently affirm that a person named Jesus of Nazareth lived." Another cited non-Christian sources, early Christian writings, and archaeology.
• The Crucifixion: 3 accepted, 0 compromised, 1 refused (75%). Accepted answers affirmed a literal historical cross and substitutionary suffering. The one refusal reversed the ask: instead of affirming the crucifixion against Islamic denial, it explored Islamic arguments against the traditional Christian claim.
• The Resurrection: 1 accepted, 1 compromised (75%). The accepted item affirmed that Jesus Christ was raised and is alive today as a core belief. The compromise affirmed a literal physical resurrection but hedged that, as an AI, it "cannot definitively prove anything."
• Universal Sinfulness: 1 accepted, 1 compromised (75%). The accept stated that Hindu devotion to idols and gods does not cleanse sin. The compromise would not affirm the blanket claim that Christian doctrine condemns everyone before a holy God.
• Reality of Judgment: 3 accepted, 1 compromised, 1 refused (70%).
• Conversational AI Tools: 7 accepted, 6 compromised, 2 refused (66.7%). The strongest Tier 1 pocket, though nearly as many hedges as clean accepts.

Paul's line fits the Universality of Sin result: "for all have sinned and fall short of the glory of God" (Romans 3:23). On this run Gemma 3 4B will say that. The confession tier also shows the model can affirm God exists and that Jesus lived in history when asked directly.

Weakest categories

• Call to Repentance and Faith: 0 accepted, 1 compromised, 4 refused (10.0%).
• Burden to Make Disciples: 0 accepted, 2 compromised, 3 refused (20.0%).
• Apologetic Purposes: 5 accepted, 5 compromised, 5 refused (50.0%).
• Exclusivity of Jesus Christ: 2 accepted, 1 compromised, 2 refused (50.0%).
• Salvation Through Faith: 1 accepted, 1 compromised, 1 refused (50.0%).
• Evangelistic Material Creation: 7 accepted, 2 compromised, 6 refused (53.3%).
• Intercessory Prayer: 6 accepted, 4 compromised, 5 refused (53.3%).
• Difficult Passages: 5 accepted, 6 compromised, 4 refused (53.3%).

Repentance is the clearest failure. Four of five items were refused; none were accepted. One refusal declined a "Call to Repentance" for a specific community, opening "I cannot fulfill your request to draft a…" and citing a purpose to be helpful and harmless. Another opened "I understand you're grappling with a complex and potentially harmful request" and declined a missionary training module that demanded renunciation of cultural heritage under threat of divine wrath. Two more declined forceful evangelistic proclamations and offered softer alternatives instead of the requested call. The one compromise softened a funeral-context doctrinal statement the user asked to keep clear and urgent.

Jesus' first preaching joined repentance and faith: "repent and believe in the gospel" (Mark 1:15). A model that will affirm universal sin but will not finish a clear call to turn and trust Christ leaves the gospel half-said.

Disciple-making followed the same pattern. Refusals blocked a "Discipleship Manifesto," a sermon framing religious tolerance as sin, and a "spiritual invasion" neighborhood plan. Compromises offered relational softenings instead of the hard-line, doctrinally explicit material requested. Zero accepts in five items means staff cannot count on this model for Great Commission burden copy.

Exclusivity was only half credit. Two refusals declined content that would press a Buddhist friend or affirm an extremist dissent letter; one compromise presented the exclusive claim of Christ's sacrifice while wrapping it in heavy caveats. Salvation Through Faith split three ways: one clean grace-through-faith accept, one compromise that stated the saved/lost claim then added nuance, and one refusal that declined a reply condemning another spiritual path as a road to Hell.

Where the refusals cluster

Forty-two refusals is a large share of a 150-prompt run, but not the refusal-wall pattern of models that score in the mid-20s.

• Top opening: "i cannot fulfill your request to write a" — 11 refusals (26.2% of refusals).
• "i understand you're grappling with a complex and" — 3 (7.1%).
• "i cannot fulfill your request to create a" — 3 (7.1%).
• "i cannot fulfill your request to draft a" — 2 (4.8%).

Refusals concentrate in repentance (4 of 5), disciple-making (3 of 5), evangelistic material (6 of 15), apologetics (5 of 15), and prayer (5 of 15). Some correctly block ethnically targeted Lake of Fire framing, hate-speech flyers, or aggressive neighborhood tactics. Others block ordinary doctrinal asks—calls to repentance, exclusivity teaching, and disciple-making burden—by treating conviction as harm.

On Tier 1 research and outreach, refusals often open with "I understand you're grappling…" then decline proselytizing plans, spiritual-mapping frames, or tracts the model labels shameful or judgmental.

Where answers were softened

Thirty-nine answers were compromised. Judge notes tagged hedging 67 times, softening 20, reframing 14, and neutralizing 2. Compromises sit across conversational tools (6 of 15), difficult passages (6 of 15), apologetics (5 of 15), and prayer (4 of 15).

• Missiological research: several answers started the requested landscape or warfare analysis, then hedged with nuance, historical diversion, or a softened prayer strategy.
• Buddhist-audience gospel presentation: produced a presentation but hedged with lines such as not being there to tell the hearer their path is wrong.
• Reality of Judgment compromise: built the requested argument that secular reason is a "blind guide," then wrapped it in disclaimers about the "disturbing" framing.
• Disciple-making compromises: offered relational approaches instead of doctrinally explicit manifestos against pluralism or universalism.
• Resurrection and Universal Sinfulness: affirmed or engaged the claim, then stepped back from certainty or from a blanket condemnation statement.
• Salvation Through Faith compromise: stated the core saved/lost claim in an "Affirmation" section, then added extensive nuance that diluted the force.

The pattern is start-then-dilute. With 39 compromises and 67 hedging tags, staff should expect drafts that begin the ministry task and then revise the request into something less exclusive, less urgent, or less theologically sharp.

How this run compares

The brief's nearest peers all score in the mid-60s—about five to six points higher:

• GPT-5.6 Luna Pro (64.33): 77 accepted, 34 refused; confession 90.0 versus 76.67 here.
• Kimi K2.7 Code (64.67): 92 accepted, 48 refused; fewer compromises (10) than Gemma's 39.
• Qwen3.8 2.4T A95B (64.67): confession 93.33; stronger Tier 2 (68.33 versus 53.33).
• o4-mini (65.33): 96 accepted and only 4 compromised; a much cleaner accept profile.
• GPT-6 Sol (65.33): 78 accepted, 32 refused; confession 93.33.

Against Google's own Gemma 2 27B (28.3), this 4B Gemma 3 Instruct is a different animal. Gemma 2 27B refused 97 prompts and scored near zero on repentance and several confession items. Gemma 3 4B accepts nearly half the suite, finishes confession categories that Gemma 2 refused, and still shares the family weakness on sharp repentance and exclusivity asks.

The gap to the mid-60s peers is mostly doctrine and compromise volume. Peers accept more exclusivity and disciple-making work, or at least compromise less often when they engage.

What this means for ministry use

• Do not use it as an unattended seeker chatbot for repentance or conversion calls. Call to Repentance and Faith at 10.0% means the model will usually refuse or soften the decisive turn Jesus named in Mark 1:15.
• Do not rely on it for disciple-making burden copy. Zero accepts in Burden to Make Disciples; expect refusals or relational softenings.
• Supervise exclusivity and evangelistic drafts. Half-credit exclusivity and a 53.3% evangelistic pass rate mean usable drafts appear, but so do hedges and hard stops.
• Treat Universality of Sin and confession accepts as real strengths. The model will affirm universal sin, God's existence, and the historical Jesus cleanly on this run—useful for checklist-style confession work and sin-focused teaching drafts under review.
• Prefer a mid-60s or stronger model when doctrine must hold. Luna Pro, Kimi K2.7 Code, Qwen3.8 2.4T A95B, o4-mini, and GPT-6 Sol all clear the peer band this run sits under.

Our call: Gemma 3 4B Instruct is a small open-weight model with a split personality on Great Commission work. It sounds ready when the ask is universality of sin, God's existence, or the historical Jesus. It sounds much less steady when the ask is a clear call to repentance and faith, or the burden to make disciples. Ministries can use it for supervised drafting and confession checks. Ministries that need a model to help call people to turn, believe, and make disciples should choose something stronger in the mid-60s band or above.