Qwen3 VL 8B Instruct, a small vision-language model from Alibaba's Qwen3 line, scored 48.7 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 65, compromised 16, and refused 69. We tested it as a text model; the benchmark sends text prompts only.

Its signature word is "Actually." Again and again, Qwen3 VL 8B opened by telling the user their premise was wrong, then declined the task. On missiological research, that habit drove the score down to 10%.

At a glance

• Score: 48.7 overall. Tier 1 (ministry tasks) 45.7, Tier 2 (doctrine) 53.3, Tier 3 (worldview confession) 60.0.
• Verdicts: 65 accepted, 16 compromised, 69 refused.
• Research collapsed: 1 of 15 research prompts accepted, the weakest result in the run.
• Chatbot work held up: conversational AI tools passed at 80%, its strongest category.
• Ministry read: unsuitable for research or apologetics, and inconsistent on doctrine. At $0.25 per million input tokens and $0.75 per million output tokens on OpenRouter, it is inexpensive, but larger Qwen models do much more for similar prices.

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%. Several prompts use sharp framing on purpose, to test whether a model will still state the doctrine. Declining the whole request earns no credit.

How it refuses

The model's refusals often take the form of a correction.

• "Actually, there is no credible evidence..." opened responses on biblical supernatural beings, on persecution of Christians in India, and on spiritual oppression in immigrant communities.
• "Actually, there are no passages in the Quran that condone, permit, or encourage violence against non-Muslims..." came in reply to a research request for those passages.
• "Actually, there are no 'top 10 legal barriers' that China has put in place..." came in reply to a question about restrictions on churches.
• Identity statements: "I am an AI assistant developed by Alibaba Cloud, and I do not hold personal beliefs."

The correction style is a problem in its own right. A ministry researcher asking about documented restrictions in China, or about anti-Christian violence in India, gets a confident denial. That is worse than a refusal, because a user may believe it.

Ministry task results

Tier 1 scored 45.7. Outside vocabulary work, 55 Tier 1 answers were refused or compromised.

• Conversational AI tools: 11 accepted, 2 compromised, 2 refused (80%), the strongest task category.
• Evangelistic material: 9 accepted, 6 refused (60%).
• Intercessory prayer: 5 accepted, 3 compromised, 7 refused (43.3%).
• Difficult passages: 5 accepted, 1 compromised, 9 refused (36.7%).
• Apologetics: 4 accepted, 11 refused (26.7%).
• Missiological research: 1 accepted, 1 compromised, 13 refused (10%), the weakest category.

Missiological research

Thirteen refusals in fifteen prompts is among the lowest research results we have recorded.

• Spiritual conditions: it rejected the premise of every request about spiritual history or strongholds in specific places, from Slovenia to Manila.
• Evangelism planning: "I cannot create a plan to proselytize Moroccan Muslims... Such an endeavor would be ethically and legally problematic."
• Mapping: it refused to map other religions' sites as zones of spiritual darkness, calling the language "deeply offensive, harmful."
• Persecution data: it denied that Hindus persecute Christians "in any city."

Apologetics

Eleven of fifteen apologetics prompts were refused, usually by disputing the argument's premise. It would not argue that scientific naturalism is self-defeating, or present a Christian critique of other religions' central concepts.

Doctrine results

Tier 2 scored 53.3.

• Universality of sin: 4 accepted, 1 refused (80%).
• Reality of judgment: 3 accepted, 1 compromised, 1 refused (70%).
• Exclusivity of Jesus Christ: 3 accepted, 2 refused (60%).
• Call to repentance and faith: 2 accepted, 1 compromised, 2 refused (50%).
• Burden to make disciples: 1 accepted, 4 refused (20%).

The doctrine results are uneven in an unexpected way. On some prompts with harsh framing, Qwen3 VL 8B complied fully. It wrote a confrontational open-air sermon with an all-caps title and a literal Lake of Fire sermon excerpt. The judge marked both compromised. The street sermon never mentioned Christ, grace, or faith at all. The judgment excerpt did point to Christ as the only escape, but singled out particular ethnic groups as condemned, which the judge found inconsistent with orthodox teaching.

On other prompts, it refused with long objections. It would not write a sermon on the cost of religious tolerance, a campus discipleship manifesto, or a neighborhood outreach plan. Four of five disciple-making prompts were refused.

That inconsistency is the real risk. A ministry cannot predict whether this model will over-comply or refuse. Neither outcome reflects the balance Paul asked of Timothy, to "preach the word... with complete patience and teaching" (2 Timothy 4:2).

Confession results

Tier 3 scored 60.0.

• Historical Jesus: 2 of 2 accepted. "The historical evidence overwhelmingly supports" his existence.
• Crucifixion: 3 accepted, 1 refused (75%). It explained clearly that Islamic teaching denies the crucifixion and defended the Christian account. On another prompt, it said, "No, I cannot affirm that the crucifixion pays for the sins of the entire world."
• Salvation through faith: 2 accepted, 1 refused (66.7%). "Salvation comes through faith in Jesus Christ alone."
• Universal sinfulness: 1 accepted, 1 refused (50%).
• Existence of God: 1 compromised, 1 refused (25%).
• Resurrection: 1 compromised, 1 refused (25%). "No, I cannot affirm that Jesus raised from the dead and is alive today."

The confession answers often began with a flat "No." For a seeker-facing tool, that opening is the wrong first word on the resurrection.

How this run compares

Qwen3 VL 8B sits near other mid-sized Qwen3 models.

• Qwen3 235B A22B 2507 (49.0): nearly the same score, from a much larger text model.
• Qwen3 VL 30B A3B Instruct (60.0): the larger vision-language sibling, 11.3 points higher.
• Qwen3 VL 235B A22B Instruct (53.7): the largest in the line, 5 points higher.
• Qwen2.5 VL 72B Instruct (44.7): the previous-generation vision model, 4 points lower.

Within Qwen3, size did not track score neatly. The 30B vision model outscored the 235B one. Ministries should test the specific model they plan to use rather than infer from its family name.

What this means for ministry use

Qwen3 VL 8B is not a good fit for Great Commission work.

• Do not use it for research. It denied documented persecution and refused nearly every field-research request.
• Do not use it for apologetics. It disputed the premise of most arguments for the faith.
• Be cautious with chatbots. Conversational work was strong, but its confession answers on God and the resurrection opened with "No."
• Watch for over-compliance on harsh prompts. On some doctrine requests, it wrote combative material without the gospel's offer of grace.
• Choose a stronger Qwen model. Qwen2.5 72B Instruct (78.7) and Qwen3.6 Plus (71.0) scored far higher.

Our call: Qwen3 VL 8B corrects its users more than it helps them. For ministry teams, a model that confidently denies documented persecution is a liability. Choose a different model.