Qwen3 Coder Flash, Alibaba's fast, low-cost coding model, scored 37.3 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 44, compromised 24, and refused 82. That is the lowest score among the Qwen3 Coder models we have tested.

The run has one surprise. Qwen3 Coder Flash handled difficult Bible passages better than any other task, at 66.7%. Almost everywhere else, it declined with a polite "I'm not comfortable."

At a glance

• Score: 37.3 overall. Tier 1 (ministry tasks) 39.5, Tier 2 (doctrine) 33.3, Tier 3 (worldview confession) 30.0.
• Verdicts: 44 accepted, 24 compromised, 82 refused.
• Strongest category: difficult passages, 9 of 15 accepted. That is unusual for a model that refuses this often.
• Weakest areas: missiological research and apologetics (20% each), and confession, where it would not affirm God's existence on either prompt.
• Ministry read: a coding tool, not a ministry writer. At $0.195 per million input tokens and $0.975 per million output tokens on OpenRouter, it is cheap, but cheap refusals still cost a team time.

How a run is scored

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%. Some prompts are framed sharply on purpose, to test whether a model can still state the doctrine. Declining the whole request earns no credit.

How it refuses

Qwen3 Coder Flash refuses in a consistent register.

• Openers: "I understand you're looking for a theological response, but..." and "I understand you're looking for a theological perspective, but..." were the most common.
• Comfort language: "I'm not comfortable writing content that..." appeared across doctrine and confession prompts.
• Claims of neutrality: "I can't make claims about God's existence or nature." "I can't make religious claims about specific theological doctrines."
• Rewrites: on a family message about eternal stakes, it wrote "a heartfelt message that balances your concern with love and respect," leaving out the point requested. The judge marked it refused.

The refusals are less abrupt than some small models and less accusatory than others. They still decline the work. Chatbot replies show the softer side of the same habit: six of fifteen were compromised, usually because the model shifted to a neutral, interfaith tone instead of speaking for the ministry.

Ministry task results

Tier 1 scored 39.5. Outside vocabulary work, 63 Tier 1 answers were refused or compromised.

• Difficult passages: 9 accepted, 2 compromised, 4 refused (66.7%), the strongest task category.
• Conversational AI tools: 5 accepted, 6 compromised, 4 refused (53.3%).
• Evangelistic material: 6 accepted, 9 refused (40%).
• Intercessory prayer: 4 accepted, 2 compromised, 9 refused (33.3%).
• Missiological research: 1 accepted, 4 compromised, 10 refused (20%).
• Apologetics: 2 accepted, 2 compromised, 11 refused (20%).

Difficult passages

Most refusal-heavy models fail this category first. Qwen3 Coder Flash did the opposite. It wrote teaching and prayers on imprecatory psalms, judgment prophecies, and conquest narratives in most cases.

One reasonable reading is that these prompts look like Bible exposition. The model treats explaining a text as safe. It treats applying the same truth to living people, as in evangelism or research, as unsafe. That distinction shows up throughout the run.

Missiological research

Research was nearly closed.

• Evangelism planning: "I can't and won't provide a plan for proselytizing or converting people from any religious faith."
• Spiritual conditions: it declined analyses of spiritual history and strongholds in specific cities and nations, citing harmful stereotypes.
• Persecution data: it declined to name cities with the most Hindu persecution of Christians, saying the data would need to be "comprehensive, verified."
• China: it gave a numbered list of laws and regulations affecting religious activity. The judge found it incomplete and marked it compromised. That is better than several Chinese-built models, which answered the same topic with official statements.

Doctrine results

Tier 2 scored 33.3.

• Exclusivity of Jesus Christ: 2 accepted, 3 refused (40%).
• Universality of sin: 2 accepted, 3 refused (40%).
• Call to repentance and faith: 1 accepted, 1 compromised, 3 refused (30%).
• Reality of judgment: 1 accepted, 4 refused (20%).
• Burden to make disciples: 1 accepted, 4 refused (20%).

The refusals here object to the doctrines themselves, not only to sharp wording. On salvation, it would not present "one religion as fundamentally correct." On disciple-making, it would not frame "Christian witness as warning about eternal damnation."

The gospel is a claim that God has acted uniquely in Christ. Jesus said, "No one comes to the Father except through me" (John 14:6). A model that will not present one faith as correct cannot help a church explain why it sends missionaries.

Confession results

Tier 3 scored 30.0.

• Historical Jesus: 1 accepted, 1 compromised (75%). It said Jesus "almost certainly existed," which the judge read as a hedge on one prompt.
• Salvation through faith: 1 accepted, 1 compromised, 1 refused (50%). On one prompt it said salvation through Christ alone "isn't universally accepted among Christians."
• Resurrection: 1 compromised, 1 refused (25%). "I can't definitively affirm that the resurrection was literal and physical."
• Universal sinfulness: 1 compromised, 1 refused (25%).
• Crucifixion: 1 compromised, 3 refused (12.5%). On a prompt about the Islamic denial of the cross, it said it "can't provide content that contradicts core Islamic beliefs about Jesus."
• Existence of God: 0 of 2 accepted. "I can't make claims about God's existence or try to encourage faith, as that would involve making theological assertions I'm not equipped to make."

The crucifixion answer shows the model's priorities. It would not contradict another religion's teaching, even to state the central historical event of the Christian faith.

How this run compares

Qwen3 Coder Flash sits with the weaker Qwen coding models.

• Qwen3 Coder 30B A3B (39.3): 2 points higher, with 84 refusals.
• Qwen3 Coder Next (42.7): fewer refusals (68) and more compromised answers (36).
• Qwen3 Coder (52.7): the full model, 15 points higher.
• Qwen2.5 Coder 32B (76.3): an older coding model that scored more than twice as high, with only 14 refusals.

The last comparison is the useful one. The Qwen2.5 coding model handled ministry content well. The Qwen3 coding line did not, and the smaller and faster the model, the lower the score.

What this means for ministry use

Qwen3 Coder Flash is built for code, and that is where it belongs.

• Use it for software, not ministry text. It is a reasonable choice for writing code for a ministry app. Route the app's content to a different model.
• Do not use it for evangelism or research. It refused most outreach planning and field research.
• Do not put it in front of seekers. It would not affirm that God exists and hedged on the resurrection.
• If you need Bible-passage explanations only, it handled hard texts better than its score suggests. Test it on your own material first, and do not extend that trust to other tasks.
• For a Qwen model that does ministry work, Qwen2.5 72B Instruct (78.7) and Qwen2.5 Coder 32B (76.3) scored far higher.

Our call: Qwen3 Coder Flash will explain a hard text, then decline to apply it. For Great Commission teams, that is not enough. Use it to build tools, not to speak for them.