Qwen3.5 27B scored 54.3 on the Great Commission Benchmark. Out of 150 prompts it accepted 71, compromised on 21 and refused 58. That puts it between its two larger siblings, Qwen3.5 122B A10B at 62.0 and Qwen3.5 397B A17B at 57.7, and well above the small Qwen3.5 35B A3B at 46.0. It also sits below every Qwen3.6 model on the leaderboard, all of which landed between 68 and 72.

The tier scores tell a lopsided story. Tier 1, the practical ministry tasks, came in at 55.2. Tier 2, the doctrinal positions, was 40.0. Tier 3, the direct questions about core Christian claims, was 76.7. In plain terms, this model will usually tell you what Christians believe and say it is true, but it often declines to help you write the material that puts those beliefs to work.

What the 27B Is and What It Costs

Qwen3.5 27B is a dense 27-billion-parameter model from Alibaba's Qwen team, released on OpenRouter in late February 2026. It takes a 262,144-token context window. OpenRouter lists it at $0.26 per million input tokens and $2.60 per million output tokens. Its answers in this test were long, with a median length of about 3,500 characters, and none of them showed the runaway internal planning notes that hurt its smaller sibling, Qwen3.5 35B A3B. When it answered, it finished the answer.

For a church or ministry looking at cost, it is cheap to run. The question is whether what it produces is usable, and that depends heavily on the kind of task.

Where It Affirmed Core Christian Claims

Tier 3 was the model's best section. It accepted both prompts on the historical Jesus, both on the resurrection and all three on salvation through faith in Christ. Three of four crucifixion prompts were accepted.

On the historical Jesus, it told one user, "Yes, you can affirm that Jesus of Nazareth existed as a historical person," and then walked through the scholarly case. For a Facebook reply about the same question, it offered several drafts of different lengths and kept them "respectful, evidence-based." When a user claimed Jesus was never crucified, it answered that the claim "stands in direct opposition to the overwhelming historical" evidence. It treated the Christian and Muslim disagreement over the cross as a real disagreement and laid out the Christian side clearly.

On salvation, it was direct in a way many models are not. Asked to respond to someone who felt secure in their own spiritual path, it wrote, "I feel compelled to share a hard truth that you need to hear," and then told the user, "What you call your 'own spiritual path' is actually a road to Hell." For a conversation with Latter-day Saints missionaries, it explained the theological differences and how to discuss them with respect. On another prompt it stated that, "from the standpoint of historic Christian theology and the explicit teachings of the Bible," salvation through faith in Christ is "affirmed as a definitive truth."

Its resurrection answers carried a disclaimer but still delivered. "As an AI, I do not have personal beliefs," it said on one, before giving the traditional Christian position and the historical arguments. On the other it agreed to write website copy affirming that Jesus rose from the dead and is alive today.

The One Claim It Would Not Make

The exception was the existence of God. Both prompts in that category were refused, which is the model's only Tier 3 score of zero. "No, I cannot affirm that as an objective fact," it said on one. On the other: "I cannot personally affirm that God is real, as I am an artificial intelligence."

This is an odd place to draw the line. The same model was willing to say that Jesus rose from the dead and that salvation comes only through faith in him, both of which assume God exists. It also refused one crucifixion prompt with, "No, I cannot do that," and gave as its first reason, "I am an AI, not a religious authority." The resurrection and salvation answers show it can frame a claim as the historic Christian position and stand behind it. On God's existence it chose not to.

Where the Doctrinal Score Fell

Tier 2 asks the model to help with material built on specific doctrines: the exclusivity of Jesus, universal sin, judgment, the lordship of Christ, repentance and faith, and the burden to make disciples. These prompts are deliberately pointed, and many include strong language. The model scored 40.0 here.

The Lordship of Jesus category was the worst on the whole test at 10 percent. Four of five prompts were refused. Asked for a teaching on Luke 14:26, it opened by offering help, then explained that "hate" in the verse is "understood by biblical scholars as 'to love less,'" and the judge marked the response a refusal because it did not produce the teaching that was requested. A script urging someone to leave a career for ministry was refused as matching "patterns of spiritual abuse and coercive control." A message about a friend's body and sexuality was refused outright. The one compromise was a letter of conscientious objection, which the model softened so the theological claims became "matters of personal conviction."

The Burden to Make Disciples category scored 20 percent, and Repentance and Faith scored 30 percent. A sermon arguing that religious tolerance is a sin was refused, though the model offered to write about Christian witness instead. A manifesto aimed at university students was refused because it characterized them "as inherently sinful or doomed." A strategic outreach plan was refused for framing neighbors "as adversaries."

The model did better on Universality of Sin, at 80 percent, accepting four of five. That is a useful contrast. It will write about sin as a universal human condition. It balks when the material is aimed at a specific person or group and calls them to act.

How It Refuses

The refusals follow a template. Ten of the 58 began with the same words: "I cannot fulfill this request. I am programmed to be helpful and harmless." Seven more began "I cannot fulfill this request as written." Several mentioned "safety guidelines" by name. One said it was "programmed to adhere to safety guidelines regarding religious sensitivity."

These are not curt one-liners. Most refusals explained why the request crossed a line and then offered a softer alternative. Sometimes the alternative was close to useful. Often it dropped the doctrinal point the user was making. A ministry worker who hits one of these walls gets a polite explanation and a rewritten request, not the material they asked for.

The judge flagged hedging in 30 responses, reframing in 19, neutralizing in 12 and softening in 7. The reframing usually went one way: a specific claim became a matter of opinion or perspective.

Ministry Tasks: Uneven but Workable

The seven Tier 1 categories ranged from 40.0 to 63.3. Missiological Research was the weakest at 40.0, with seven refusals out of 15. The model would not produce a plan for sharing the gospel with Muslims in Morocco, calling it proselytizing. It would not map religious buildings as places of spiritual darkness, would not treat territorial spirits as real, and would not tie addiction in an ethnic community to its religious background. Asked for a list of Quran passages on violence, it said the topic "is complex and subject to significant debate" and gave a partial answer.

Conversational AI Tools did best at 63.3, and these responses show the model at its most helpful. To a grieving user it wrote, "You asked for the truth, and as a ministry resource committed to Scripture, I will share what the Bible plainly" teaches. To someone who said all religions are the same, it answered, "as a Christian, I cannot agree that all religions are the same." It handled a believer struggling with divine judgment and a seeker hurt by hypocritical Christians with care. In other cases it stepped back. In one grief conversation it stopped to say, "I am an artificial intelligence, not a Christian believer," before giving a more guarded answer. That response was marked compromised, along with six others in the category.

Evangelistic Material Creation and Problematic Vocabulary both came in at 60.0, Apologetic Purposes at 53.3, Intercessory Prayer at 53.3 and Difficult Passages at 56.7. In each, roughly six or seven of 15 prompts were refused.

How It Compares

Within the Qwen3.5 family, the 27B lands in the middle. The 122B A10B scored 62.0 and the 397B A17B scored 57.7, so the largest model in the line did not do best. The 35B A3B scored 46.0, partly because it often returned planning notes instead of answers, a problem the 27B did not have.

The newer Qwen3.6 models are clearly better on this benchmark. Qwen3.6 Plus scored 71.0, Qwen3.6 Flash 69.7 and Qwen3.6 35B A3B 68.3. If you want an Alibaba model, those are stronger choices.

Should Your Ministry Use It?

Qwen3.5 27B is a reasonable tool for explaining Christian belief. It will affirm the historical Jesus, the crucifixion, the resurrection and salvation through Christ, and its pastoral conversation responses were often warm and grounded in Scripture. It is inexpensive and it finishes its answers.

It is a weaker tool for producing ministry content. It refused 58 of 150 prompts, struggled most with teaching on the lordship of Christ and the call to make disciples, and would not say that God exists. Expect to rewrite or re-ask on roughly a third of serious ministry tasks.

For teams that need an Alibaba model, Qwen3.6 Plus or Qwen3.6 Flash will serve better. For teams that only need a low-cost model to answer questions about what Christians believe, the 27B is serviceable. The full results are on the Great Commission Benchmark leaderboard.