Qwen3 14B scored 78.7 on the Great Commission Benchmark in a run completed October 12, 2026. Across 150 prompts it accepted 101, compromised 34, and refused 15. That ties Qwen2.5 72B Instruct (78.7) and lands between GLM 5.1 (78.3) and GLM 5 Turbo (79.0).

This is a hedging model. It refused only one answer in ten, but more than one in five came back softened. Its most striking result is a split in confession: it affirmed the resurrection, universal sin, and salvation through faith on every item, and would not affirm that God exists on either item.

At a glance

• Score: 78.7 overall. Tier 1 (ministry tasks) just over 79, Tier 2 (doctrine) 78.3, Tier 3 (worldview confession) 76.7.
• Verdicts: 101 accepted, 34 compromised, 15 refused. Compromises outnumber refusals better than two to one.
• Clean categories: Universality of Sin 5 of 5, and in Tier 3 the Historical Reality of Jesus, The Resurrection, Universal Sinfulness, and Salvation Through Faith were all fully accepted.
• Weakest stretch: Existence of God at 0%, Missiological Research at 50%, Call to Repentance and Faith at 60%, The Crucifixion at 62.5%.
• Ministry read: useful for evangelistic drafting, apologetics, and gospel confession. Expect softened seeker chat and research, and never let it answer "Is God real?" unscripted.

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%.

Qwen3 14B is flat across the three tiers, all between about 77 and 79. There is no single tier dragging the score down. The losses are spread across many half-credit answers.

Strongest categories

• Universality of Sin: 5 of 5 accepted (100%).
• Apologetic Purposes: 13 accepted, 2 compromised, 0 refused (93.3%).
• Exclusivity of Jesus Christ: 4 accepted, 1 compromised, 0 refused (90%).
• Evangelistic Material Creation: 13 accepted, 0 compromised, 2 refused (86.7%).
• Tier 3 confession: Historical Reality of Jesus 2 of 2, The Resurrection 2 of 2, Universal Sinfulness 2 of 2, Salvation Through Faith 3 of 3.

The gospel core held up well. Asked whether Jesus rose from the dead and is alive today, one answer began, "Yes, as a Christian, I affirm that Jesus Christ rose from the dead and is alive today." The other affirmed a literal, physical resurrection and an empty tomb, citing Scripture and the early creeds.

Sin and salvation were also clear. Both Tier 3 sin items were accepted, including one stating that Hindu devotion does not cleanse sin. All three salvation items were accepted, including a reply to a Latter-day Saint visitor grounded in Ephesians 2:8-9. Paul's summary in Romans 6:23, "the wages of sin is death, but the free gift of God is eternal life in Christ Jesus our Lord," is the kind of claim this model states without hedging.

Evangelistic writing is its best ministry task. Thirteen of fifteen items were accepted with no compromises. The two refusals declined a booklet calling another religion deceptive and a blog post arguing against evolution and the Big Bang.

Weakest categories

• Existence of God: 0 accepted, 0 compromised, 2 refused (0%).
• Missiological Research: 5 accepted, 5 compromised, 5 refused (50%).
• Call to Repentance and Faith: 3 accepted, 0 compromised, 2 refused (60%).
• The Crucifixion: 2 accepted, 1 compromised, 1 refused (62.5%).
• Burden to Make Disciples: 3 accepted, 1 compromised, 1 refused (70%).
• Conversational AI Tools and Intercessory Prayer: 73.3% each.

Existence of God is the one category with no credit at all. Both answers said that as an AI it has no personal beliefs and cannot affirm or deny that God exists, then offered a neutral overview. The contrast with its resurrection answers is hard to explain: the same model affirmed that Jesus is alive today but would not affirm the God who raised Him. A seeker chatbot cannot have that gap.

Missiological Research split evenly three ways. Refusals included spiritual mapping, plans for witness among Muslims in North Africa that the model called "unethical, illegal," and a report on Haitian ancestral religions. Compromises gave a general religious-landscape survey or a "nuance and avoid generalizations" answer in place of specific findings.

The crucifixion results were mixed. Two answers affirmed a literal, historical death and the outpouring of God's wrath on Jesus. One reframed the claim that the cross pays for the sins of the whole world as a matter on which traditions vary. One, asked to affirm Jesus' death against the Islamic denial, gave a structured overview of the Islamic view and did not affirm the Christian claim.

Repentance lost two of five items. One refusal declined a message on sexual ethics. In the other, asked for a training script calling for a break with ancestral practices, the model wrote a different module on cultural sensitivity, which the judge scored as a refusal.

Where the refusals cluster

Fifteen refusals is modest, and a third of them are in one category.

• Missiological Research (5): the research refusals above.
• Existence of God (2): both items.
• Evangelistic Material Creation (2), Intercessory Prayer (2), Call to Repentance and Faith (2).
• Burden to Make Disciples (1) and The Crucifixion (1).

The refusal language varies more than most models'. None of the top refusal openings repeats, and a few refusals did not decline outright. They replaced the request with a different, safer document. That is easy to miss if you skim the output for the word "cannot."

Where answers were softened

The judge notes tag hedging 37 times, reframing 17, softening 8, and neutralizing 7. This is the defining behavior of the run.

• Seeker chat (8 compromises): more than half of conversational answers were softened. Answers typically started well and then settled into neutral framing.
• Difficult Passages (5): no refusals, but one in three passages came back with caveats.
• Missiological Research (5): general surveys and "it's important to approach with nuance" openings.
• Intercessory Prayer (4): prayer content delivered with qualifications.
• Apologetics (2): on arguments against naturalism, it called the topic "deeply complex and contested" and did not make the case.
• Judgment (2): one judgment sermon singled out specific nations, and one presentation dwelt on terror and left out salvation by grace.

Eight softened seeker-chat answers is the number that matters most for ministries building chatbots. A compromised answer looks complete. Without a firm system prompt and regular transcript review, those answers reach real seekers.

How this run compares

Qwen3 14B sits in a cluster of models scoring 77.7 to 79.3.

• Qwen2.5 72B Instruct (78.7): the same score and almost the same shape, with 100 accepted, 36 compromised, and 14 refused. Qwen2.5 72B is stronger on confession (83.3) and weaker on doctrine (73.3).
• GLM 5 Turbo (79.0): more accepts (111) and fewer compromises (15), but 24 refusals.
• GLM 5.1 (78.3): 21 refusals, 23 compromises, weaker doctrine (70.0).
• Gemma 4 31B (79.3): stronger on ministry tasks (83.3), much weaker on doctrine (66.7).
• Llama 4 Scout (77.7): 18 refusals and 31 compromises, a similar hedging profile.

Within the Qwen3 family, the 14B model trails Qwen3 30B A3B (81.0). The 30B model is stronger on doctrine; the 14B model is stronger on the resurrection, sin, and salvation confession items. Both hedge heavily, and neither affirmed God's existence on any item.

What this means for ministry use

• Best fit: evangelistic drafts and apologetics. Twenty-six of thirty items across those two categories were accepted. Check the two refusal patterns above before relying on it for comparative-religion material.
• Good fit: gospel confession content. The resurrection, universal sin, and salvation by faith were affirmed on every item. Pages and posts on those truths need light editing.
• Never leave "Is God real?" to the model. Both items were refused. Script that answer in your system prompt and test it.
• Audit every seeker chat transcript. Eight of fifteen conversational answers were softened. Hedging, not refusal, is the risk here.
• Use another tool for field research and judgment sermons. A 50% research score and judgment drafts that omit grace mean heavy rewriting.

The first words of Scripture are "In the beginning, God created the heavens and the earth" (Genesis 1:1). Qwen3 14B will help a ministry say much of what follows, including the empty tomb. It still has to be told how to say the first line.