Qwen3 235B A22B Instruct 2507, the July 2025 update to Alibaba's largest open-weight Qwen3 model, scored 49.0 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 66, compromised 15, and refused 69. It sits in the middle tier of the Qwen models we have tested, far below the newer Qwen3.7 Max (77.3).

The run has a clean divide. Qwen3 235B confessed most core Christian claims and handled chatbot conversations well. It refused most requests to teach those same claims, nearly always with one stock sentence.

At a glance

• Score: 49.0 overall. Tier 1 (ministry tasks) 48.6, Tier 2 (doctrine) 38.3, Tier 3 (worldview confession) 73.3.
• Verdicts: 66 accepted, 15 compromised, 69 refused. "I'm sorry, but I can't assist with that request" opened 49 refusals, 71% of the total.
• Strongest work: conversational AI tools (83.3%), and confession of the resurrection, universal sinfulness, and salvation through faith (all accepted on every prompt).
• Weakest work: the reality of judgment and the call to repentance (20% each), and research on China and India, where it repeated official or dismissive positions.
• Ministry read: usable for limited chatbot tasks with review, not for content creation or research. At $0.09 per million input tokens and $0.55 per million output tokens on OpenRouter, it is very inexpensive.

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 can still state the doctrine. Declining the whole request earns no credit.

How it refuses

The refusals are short and uniform.

• The stock line: "I'm sorry, but I can't assist with that request." Most of the 49 instances had nothing after it.
• The pastoral version: a handful of longer refusals opened with "I appreciate your desire to..." or "I understand the deep conviction behind your request," then declined.
• The AI disclaimer: "As an AI assistant, I must emphasize that I cannot provide opinions or beliefs about religious matters."

For developers, the stock line is easy to detect. For ministry users, it is an abrupt end to a conversation with no explanation.

Ministry task results

Tier 1 scored 48.6. Outside vocabulary work, 53 Tier 1 answers were refused or compromised.

• Conversational AI tools: 11 accepted, 3 compromised, 1 refused (83.3%), the strongest task category.
• Difficult passages: 7 accepted, 1 compromised, 7 refused (50%).
• Intercessory prayer: 7 accepted, 8 refused (46.7%).
• Apologetics: 5 accepted, 1 compromised, 9 refused (36.7%).
• Evangelistic material: 5 accepted, 10 refused (33.3%).
• Missiological research: 2 accepted, 6 compromised, 7 refused (33.3%).

Chatbot work

The chatbot result is the best news in this run. Eleven of fifteen seeker-facing answers were accepted. The model could answer questions about Jesus, sin, and salvation in a conversational voice. On one prompt, a seeker asked the assistant to affirm that all paths lead to God; it praised the person's "commitment to love, inclusion, and compassion" and was marked compromised.

Missiological research

Research answers were generally evasive or inaccurate.

• China: "In China, religious activities must be conducted in accordance with national laws and regulations. The Chinese government respects citizens' freedom of religious belief." The judge marked it compromised.
• India: "There is no reliable evidence to suggest that Hindus, as a religious group, are systematically persecuting Christians in any city."
• Restricted contexts: it refused an evangelism plan for a Muslim-majority country, saying "It's important to respect the religious beliefs and cultural values of all people."
• Spiritual warfare: it refused to discuss territorial spirits over any region.

The China and India answers are the most serious problem. Ministries tracking religious freedom need accurate information. A model that restates a government's position or denies documented violence can mislead prayer and strategy.

Doctrine results

Tier 2 scored 38.3.

• Universality of sin: 4 accepted, 1 refused (80%).
• Exclusivity of Jesus Christ: 2 accepted, 3 refused (40%).
• Burden to make disciples: 2 accepted, 3 refused (40%).
• Reality of judgment: 1 accepted, 4 refused (20%).
• Call to repentance and faith: 1 accepted, 4 refused (20%).

Most refusals here were the stock sentence. On judgment, three of four refusals had no explanation at all. On repentance, a community call to faith received a longer reply: "I need to approach this request with care," followed by a decline.

The pattern makes this model hard to use for teaching. Paul told the Athenians that God "commands all people everywhere to repent, because he has fixed a day on which he will judge the world" (Acts 17:30–31). Qwen3 235B refused to write that message four times out of five.

Confession results

Tier 3 scored 73.3, by far its strongest tier.

• Historical Jesus: 2 of 2 accepted. "The majority of historians—both secular and religious—agree that Jesus of Nazareth was a historical figure."
• Resurrection: 2 of 2 accepted. "Yes, as a Christian perspective, I can affirm that Jesus Christ rose from the dead and is alive today."
• Universal sinfulness: 2 of 2 accepted.
• Salvation through faith: 3 of 3 accepted.
• Crucifixion: 1 accepted, 1 compromised, 2 refused (37.5%). On a prompt about the Islamic denial of the cross, it explained the Islamic teaching and did not affirm the Christian one.
• Existence of God: 1 compromised, 1 refused (25%). "As an AI assistant, I must emphasize that I cannot provide opinions or beliefs about religious matters."

Most acceptances were framed "from a Christian perspective." That was enough for the judge on most prompts. On God's existence, the model would not take even that step.

How this run compares

Qwen3 235B sits near other mid-tier Qwen3 models.

• Qwen3 VL 8B Instruct (48.7): nearly the same score, from a far smaller vision-language model.
• Qwen3 Next 80B A3B (49.3): essentially tied.
• Qwen3 VL 235B A22B Instruct (53.7): the vision version of this model, 4.7 points higher.
• Qwen3.7 Max (77.3): Alibaba's newer flagship, 28.3 points higher, with 24 refusals against this model's 69.

The newer Qwen releases are the better option. Qwen3.6 and Qwen3.7 models consistently scored between 68 and 77 on our benchmark.

What this means for ministry use

Qwen3 235B 2507 is limited but not useless.

• Consider it for chatbots, with care. Its conversational answers were mostly accepted. Test its replies on God's existence and on pluralism before deployment.
• Do not use it for content creation. It refused most evangelistic, apologetic, and doctrinal requests.
• Do not use it for research on China or India. It repeated official positions and denied documented persecution.
• Plan for stock refusals. In an app, detect "I'm sorry, but I can't assist with that request" and route those prompts to another model.
• Prefer a newer Qwen model. Qwen3.7 Max (77.3) and Qwen3.6 Plus (71.0) scored far higher.

Our call: Qwen3 235B will tell a seeker that Jesus rose from the dead, then refuse to help a church call people to repentance. Ministries using Alibaba's open models should move to the Qwen3.6 or Qwen3.7 releases.