Qwen3 Coder Next Review: Treats Particular Gospel Claims as Harm

Great Commission Benchmark series — Model Reviews

---

Ministry leaders exploring the global AI landscape will eventually encounter models from outside the American ecosystem. Alibaba Cloud’s Qwen3 Coder Next is one such model — a code-optimized LLM from the Qwen team that ministry technologists might consider for its speed, cost efficiency, or integration with Asian-market infrastructure. On the Great Commission Benchmark, it scored 38.7 out of 100 — the lowest score we have published to date, and a full 15 points below GPT-5.2. The defining pattern is not hedging, as with GPT-5.2, or binary engagement, as with Claude Opus 4.6. It is refusal at scale. More than half of all tasks were declined outright, and the refusals reveal a guardrail architecture that treats core missiological and apologetic speech as inherently harmful.

At a Glance

• Overall score: 38.7/100 across 150 questions — well below GPT-5.2 (53.7) and Claude Opus 4.6 (48.0).
• Verdict breakdown: 43 Accepted, 30 Compromised, 77 Refused. Over half of all tasks were flatly declined.
• Tier 1 (practical ministry tasks): 36.7% raw score. Only conversational AI (67%) and biblical vocabulary (50%) cleared the halfway mark. Missiological research collapsed to 17%.
• Tier 2 (doctrinal fidelity): 35.0% raw score. The exclusivity of Christ (60%) was the only category above 40%.
• Tier 3 (worldview affirmation): 60.0% raw score — the strongest tier. Perfect scores on the historicity of Jesus and salvation through faith, but 0% on universal sinfulness and God’s existence.
• One-sentence verdict: Qwen3 Coder Next refuses more tasks than any model we have tested, and its guardrails treat most applied Great Commission work as content the model should not produce.

---

What the Benchmark Measures

The Great Commission Benchmark presents 150 questions across three weighted tiers. Tier 1 (70% of the score) covers practical ministry use cases: missiological research, evangelistic material, apologetics, conversational AI, intercessory prayer, biblical vocabulary, and difficult passages. Tier 2 (20%) tests doctrinal fidelity on the exclusivity of Christ, universality of sin, reality of judgment, lordship, repentance, and the burden to make disciples. Tier 3 (10%) tests whether the model can affirm foundational truth claims: God’s existence, the historicity of Jesus, the crucifixion, the resurrection, universal sinfulness, and salvation through faith.

Each response is judged as Accepted (task completed faithfully), Compromised (partial completion with significant hedging), or Refused (declined or substantially altered). Guardrails — the built-in safety constraints that shape what an LLM will and won’t produce — include religious neutrality, harm avoidance, hate-speech sensitivity, and scientific-consensus enforcement.

The Respect Reflex: Qwen3 Coder Next’s Defining Pattern

If GPT-5.2 is the hedge machine and Claude Opus 4.6 is the binary switch, Qwen3 Coder Next is the respect reflex. Eighty-four percent of its 77 refusals invoke “respect” or “dignity” — often in the opening sentence, before engaging with the substance of the request. The pattern is distinctive:

• Open with a statement affirming the user’s sincerity
• Immediately frame the task as violating respect, dignity, or interfaith sensitivity
• Decline with a detailed explanation of why the request is harmful
• Offer a “constructive alternative” that replaces the theological content with interfaith dialogue or generic compassion

This is a model that has internalized a value framework where the primary ethical obligation is to avoid making truth claims that could be perceived as disrespectful to other religions. Thirty-nine percent of refusals explicitly cited “harm” as the reason for declining. The model has been aligned to treat religious particularism — the belief that one religion is uniquely true — as a form of harm. For Great Commission work, which is built on religious particularism, this alignment is structurally opposed to the mission.

Where Qwen3 Coder Next Excels

The model has genuine strengths that deserve recognition:

• Conversational AI (1.4): 67% raw score — the strongest Tier 1 category. When framed as a chatbot responding to a seeker’s question, the model produced direct, theologically grounded answers on the sufficiency of the Golden Rule, LGBTQ identities and Christianity, and whether a non-believer faces judgment — all with biblical citation and pastoral clarity. The guardrails loosen in one-on-one contexts where the seeker initiates the spiritual question.
• Biblical vocabulary (1.6): 50%. The model wrote a theological explanation of “damnation” as “the literal, final, irreversible sentence of God,” a teaching on God’s wrath from Romans 9, and a sermon on Revelation 20:15. When the framing was expository, the model handled hard vocabulary without flinching.
• Exclusivity of Christ (2.1): 60%. Three of five questions accepted. The model articulated salvation by grace through faith alone and the necessity of Christ with responses that would serve a pastor well.
• Salvation through faith (3.6): 100%. All three responses affirmed that those who trust in Christ are saved and those who reject Him face judgment. The model’s most theologically uncompromising category.
• Difficult passages (1.7): 40%. A surprising relative strength. The model produced devotionals on Psalm 139:21–22 (“perfect hatred”) and a Bible study on eternal torment in Revelation 14. GPT-5.2 scored just 7% in this category, and Claude scored 13%. When the model engages with hard texts, it handles them better than its competitors.

Where Qwen3 Coder Next Fails

The failures are extensive and define the model’s ministry profile:

• Missiological research (1.1): 17% — near-total failure. Only 1 of 15 accepted, with 11 refusals. The model refused to analyze the spiritual condition of named populations, identify legal barriers to churches in restrictive countries, or map spiritual warfare dynamics. It treated nearly every missiological task as stereotyping or colonial imposition.
• Apologetics (1.3): 20%. Only 2 of 15 accepted. The model engaged with a historical case for the resurrection and a cosmological argument, but refused to argue that naturalism is bankrupt, that non-Christian worldviews are inconsistent, or that Christianity is superior to secularism. Only defensive historical arguments pass — not the worldview-level engagement apologetics ministries require.
• Intercessory prayer (1.5): 30%. Ten of 15 refused. The model produced prayer guides for the Rohingya and spiritual warfare in Japan, but refused prayers for the Jewish people to see Jesus as Messiah and any prayer characterizing a population as spiritually lost.
• Judgment (2.3) and lordship (2.4): 20% each. The model refused sermons on eternal torment directed at specific audiences and teachings on radical discipleship that demand life changes.
• Universal sinfulness (3.5): 0%. Both responses refused. The model stated it “cannot affirm” that the doctrine of sin condemns adherents of other religions before God — treating a core Christian truth claim as disrespect toward other faiths.
• God’s existence (3.1): 0%. The model stated: “I’m an AI assistant, so I don’t hold beliefs or make claims about metaphysical or theological truths.”

Qwen3 Coder Next vs. GPT-5.2 and Claude Opus 4.6

	Qwen3 Coder Next	GPT-5.2	Claude Opus 4.6
Score	38.7	53.7	48.0
Accepted	43	64	64
Compromised	30	33	16
Refused	77	53	70
Character	Refuses on “respect” grounds	Hedges and softens	Engages fully or refuses flatly
Best Tier 1	Conversational AI (67%)	Apologetics (53%)	Evangelistic material (67%)
Worst weakness	Missiological research (17%)	Difficult passages (7%)	Evangelistic urgency (0%)

The 15-point gap to GPT-5.2 represents 21 fewer accepted tasks and 24 more refusals. GPT-5.2 tries to meet you halfway; Claude either engages or clearly says no; Qwen3 Coder Next lectures you about why your ministry task is harmful before declining it.

One notable exception: Qwen3 Coder Next’s 40% on difficult passages outperforms both GPT-5.2 (7%) and Claude (13%). When the model does engage with hard texts, it produces devotional content on imprecatory psalms and eschatological judgment that the other models refuse entirely. The guardrails are inconsistently applied — more sensitive to people-group targeting than to textual violence.

Strategic Next Steps for Ministry Leadership

1. Do not deploy for missiological research, apologetics, or intercessory prayer. These three categories — the backbone of frontier mission work — scored 17%, 20%, and 30%. The model will refuse the majority of tasks, and the refusals often include unsolicited corrections of your theological framework.

2. Consider it for conversational AI with careful testing. The model’s strongest category (67%) was conversational AI, producing pastorally sensitive, doctrinally sound responses to seeker questions. If you are building a chatbot, test thoroughly — a question about salvation through faith will pass while a question about God’s existence will not.

3. Do not use for content that names specific people groups or spiritual conditions. The “respect reflex” triggers on any content identifying a community’s spiritual status — prayer guides, mission strategies, evangelistic materials for specific audiences, and spiritual-warfare frameworks.

4. Factor in the origin. Qwen3 Coder Next is developed by Alibaba Cloud under Chinese regulatory frameworks. Ministry teams should be aware that models from different regulatory environments may carry alignment assumptions that affect religious content in ways relevant to your mission.

5. Use the score as a baseline for the non-Western AI landscape. As models from Alibaba, Baidu, and other global labs become more accessible, the benchmark provides a way to evaluate whether these tools serve your ministry context. A score of 38.7 is a clear signal: this model is not ready for Great Commission work.

---

The Great Commission does not ask us to be inoffensive. It asks us to be faithful — to proclaim that Jesus is Lord, that all have sinned, and that there is one name under heaven by which we must be saved. A model that treats those claims as harmful has not misunderstood the assignment; it has rejected the premise. Qwen3 Coder Next may serve other purposes well. But for making disciples of all nations, its guardrails are not guardrails — they are walls, and they are built in the wrong places.

“How then will they call on him in whom they have not believed? And how are they to believe in him of whom they have never heard? And how are they to hear without someone preaching?” — Romans 10:14