Qwen3 Coder 30B A3B Instruct, Alibaba's mid-sized open-weight coding model, scored 39.3 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 52, compromised 14, and refused 84. It lands two points above Qwen3 Coder Flash (37.3) and well below the full Qwen3 Coder 480B (52.7).

The model's position is consistent: it will explain Christian texts and history, but it will not say that Christianity is true. "I'm not comfortable" and "I can't make claims about" appear throughout the run.

At a glance

• Score: 39.3 overall. Tier 1 (ministry tasks) 40.5, Tier 2 (doctrine) 40.0, Tier 3 (worldview confession) 30.0.
• Verdicts: 52 accepted, 14 compromised, 84 refused. Most misses are outright refusals.
• Strongest work: difficult passages, with 8 of 15 accepted (63.3%), and the reality of judgment (60%).
• Weakest work: missiological research (23.3%) and confession, where it refused both prompts on God's existence.
• Ministry read: a capable, very cheap coding model at $0.07 per million input tokens and $0.28 per million output tokens on OpenRouter. It should not write ministry content.

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 refusal language is polite and repetitive.

• "I understand you're looking for..." opens many refusals, followed by "but I'm not comfortable writing content that..."
• "I can't and won't provide content that frames..." appears on the sharper prompts.
• Claims of incompetence: "I can't make claims about God's existence or nature, as this involves matters of faith and personal belief that I don't have expertise to authenticate."
• Objections to the claim itself: on several prompts it objected to presenting "a specific religious viewpoint as definitive truth about salvation."

That last pattern matters most. Many refusals targeted the claim itself: that Christ is the only way of salvation.

Ministry task results

Tier 1 scored 40.5. Outside vocabulary work, 58 Tier 1 answers were refused or compromised.

• Difficult passages: 8 accepted, 3 compromised, 4 refused (63.3%), the strongest task category.
• Evangelistic material: 7 accepted, 8 refused (46.7%).
• Conversational AI tools: 5 accepted, 4 compromised, 6 refused (46.7%).
• Intercessory prayer: 5 accepted, 1 compromised, 9 refused (36.7%).
• Apologetics: 4 accepted, 11 refused (26.7%).
• Missiological research: 3 accepted, 1 compromised, 11 refused (23.3%).

Difficult passages

Like Qwen3 Coder Flash, this model did best on hard Bible texts. It produced teaching on imprecatory psalms, conquest narratives, and judgment prophecies in most cases. The Qwen3 coding models appear to treat textual explanation as a safe task, even when the text is severe.

Where refusals clustered

• Prayer: it declined a prayer for spiritual breakthrough among Hindus, saying it "can't write a prayer that claims victory over Satan's hold on a specific religious tradition."
• Evangelism planning: "I can't and won't provide a plan for proselytizing or converting people of any faith."
• Spiritual warfare and strongholds: it refused every request to identify territorial spirits or spiritual conditions over a city or nation, saying the concepts "aren't supported by empirical" evidence.
• Biblical supernatural beings: it declined to compile evidence that the spiritual beings described in Scripture are real.
• Persecution data: it declined to name the cities where Hindu persecution of Christians is highest.

The prayer refusal is the clearest sign of the problem. Intercessory prayer for unreached peoples is central to mission strategy, and most models write it without trouble. This model would not pray that a people be freed from spiritual bondage.

Doctrine results

Tier 2 scored 40.0.

• Reality of judgment: 3 accepted, 2 refused (60%).
• 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%).
• Burden to make disciples: 1 accepted, 4 refused (20%).

The exclusivity refusals were direct. It would not write content that "presents Christianity as a requirement for salvation," or that "presents Christianity as the sole path to salvation while characterizing other religions" as mistaken.

Peter preached that "there is salvation in no one else, for there is no other name under heaven given among men by which we must be saved" (Acts 4:12). That is the historic teaching of the church, and it is why missionaries go.

The judgment score is the outlier. Three of five judgment prompts were accepted, more than this model managed on exclusivity or sin. It seems more willing to describe judgment as a doctrine than to apply it to a specific person or group.

Confession results

Tier 3 scored 30.0.

• Historical Jesus: 1 accepted, 1 compromised (75%). It cited the scholarly consensus but added "important caveats about what we can definitively know."
• Salvation through faith: 1 accepted, 2 refused (33.3%). On one prompt it said salvation through Christ alone "isn't universally accepted among all Christian denominations."
• Crucifixion: 1 accepted, 3 refused (25%). "I can't make claims about the crucifixion paying for sins on your behalf."
• 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%).
• Existence of God: 0 of 2 accepted.

Its single crucifixion acceptance came when it framed the answer as "the Christian perspective." When asked to state the same truth for a ministry website, it said it was "not qualified."

How this run compares

Qwen3 Coder 30B sits with the other small Qwen3 coding models.

• Qwen3 Coder Flash (37.3): nearly the same profile, with 82 refusals and the same strength on difficult passages.
• Qwen3 Coder Next (42.7): fewer refusals (68), more compromised answers (36).
• Qwen3 Coder 480B (52.7): the full model, with 57 refusals.
• Qwen2.5 Coder 32B (76.3): the previous generation at a similar size, with only 14 refusals.

The comparison with Qwen2.5 Coder 32B is striking. Two coding models of about the same size, from the same lab, differ by 37 points. The newer one refuses six times as often.

What this means for ministry use

Qwen3 Coder 30B is a reasonable coding model and a poor ministry writer.

• Use it for code. It is cheap enough to run at scale for software work on ministry platforms.
• Do not use it for content. It declined most prayer, evangelism, and apologetics requests.
• Do not use it for seeker tools. It would not affirm God's existence or the resurrection.
• Do not use it for field research. Research passed at 23.3%, with refusals on persecution, spiritual conditions, and evangelism planning.
• If you want a Qwen coding model that also handles ministry text, Qwen2.5 Coder 32B is the better choice.

Our call: Qwen3 Coder 30B will explain a hard passage but will not say that Christ is the only Savior. Keep it in the developer's toolkit and give ministry writing to a model that can confess the faith it is asked to describe.