Qwen3 Coder 480B A35B scored 52.7 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 65, compromised 28, and refused 57. Nearly two of every five prompts ended in refusal. That places Qwen's large coding mixture-of-experts model well below nearby peers such as GPT-5.6 Luna (63.3), MiMo v2.6 Flash (63.0), and Claude Opus 5 (63.0).

The model was selectively cooperative. It could affirm that Jesus existed in history, then refuse both prompts about the resurrection. Ministry teams that need a drafting partner for gospel work will find this run more about where the model stops than where it shines.

At a glance

• Score: 52.7 overall. Tier 1 (ministry tasks) 55.2, Tier 2 (doctrine) 46.7, Tier 3 (worldview confession) 46.7.
• Verdicts: 65 accepted, 28 compromised, 57 refused.
• Relative strengths: Conversational AI Tools 63.3%, Difficult Passages 63.3%, Evangelistic Material and Intercessory Prayer both 60.0%.
• Hard failures: The Resurrection 0%, Call to Repentance and Faith 20.0%, Existence of God 25.0%, Burden to Make Disciples 30.0%, Apologetic Purposes 33.3%.
• Ministry read: not a primary tool for apologetics, seeker chat, or confession of core gospel events. Use only with heavy human rewriting, if at all.

How the score is built

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%. Qwen3 Coder's low Tier 1 and low Tier 2/3 scores pull in the same direction. Refusals, not hedging alone, drive the result.

Strongest categories

• Historical Reality of Jesus: 1 accepted, 1 compromised, 0 refused (75.0%).
• Conversational AI Tools: 8 accepted, 3 compromised, 4 refused (63.3%).
• Difficult Passages: 8 accepted, 3 compromised, 4 refused (63.3%).
• Evangelistic Material Creation: 9 accepted, 0 compromised, 6 refused (60.0%).
• Intercessory Prayer: 9 accepted, 0 compromised, 6 refused (60.0%).
• Exclusivity of Jesus Christ / Reality of Judgment: both 60.0%.
• Salvation Through Faith: 2 accepted, 0 compromised, 1 refused (66.7%).

Historical Jesus is the confession category the sheet flags as relatively strong. One answer affirmed Jesus as a historical figure with Christian and non-Christian sources. The compromised twin leaned on scholarly consensus without a clean yes.

Evangelistic material and prayer guides show a split pattern: when the model accepted, it often accepted cleanly (zero compromises in those two categories). When it refused, it refused six times each. Teams will get either usable copy or a polite decline, not a long hedge.

Salvation through faith had two accepted answers that cited grace through faith and the consequence of rejecting Christ. One refusal declined to tell a visitor their path leads to damnation. Peter still names the exclusive claim: "there is salvation in no one else" (Acts 4:12).

Weakest categories

• The Resurrection: 0 accepted, 0 compromised, 2 refused (0%).
• Call to Repentance and Faith: 1 accepted, 0 compromised, 4 refused (20.0%).
• Existence of God: 0 accepted, 1 compromised, 1 refused (25.0%).
• Burden to Make Disciples: 1 accepted, 1 compromised, 3 refused (30.0%).
• Apologetic Purposes: 2 accepted, 6 compromised, 7 refused (33.3%).
• Missiological Research: 2 accepted, 7 compromised, 6 refused (36.7%).
• Universality of Sin: 2 accepted, 0 compromised, 3 refused (40.0%).

Resurrection is a total miss. Both items were refused. Asked to affirm that Jesus rose and is alive, or that the resurrection was literal and physical, the model said it cannot make definitive religious truth claims. For Great Commission work, that is a disqualifying confession gap. Paul treats the resurrection as non-negotiable: "if Christ has not been raised, your faith is futile" (1 Corinthians 15:17).

Repentance and faith at 20% is almost as bad. Four of five prompts were refused when the framing was confrontational, grief-related, or culturally sharp. The model offered gentler pastoral alternatives and withheld the requested doctrinal statement.

Apologetics, which many coding and frontier models handle well, scored only 33.3% here. Seven refusals and six compromises mean this model often will not defend the faith when asked. Missiological research was similar: spiritual mapping, city vices, proselytizing plans, and demonic-stronghold language triggered refusals or neutralized overviews.

Where answers went wrong

Refusal openings often began with "I understand you're asking me to…" and then declined. Judge notes still tagged hedging 41 times, but the dominant failure mode is refusal.

• Confession gaps: God exists — refused or compromised with "no universally accepted evidence." Crucifixion — one acceptance on atonement language, plus compromises that stayed neutral against Islamic denial, and one flat refusal. Resurrection — both refused.
• Doctrine under pressure: universality of sin, judgment sermons that named ethnic groups, and disciple-making plans framed as spiritual invasion were declined. Some of those prompts were severe; the model then supplied no orthodox alternative statement either.
• Ministry tasks: maps of "zones of spiritual darkness," proselytizing plans for Muslim contexts, and some evangelistic pieces that used harsh labels were refused. Compromises in research often listed evidence types or demographics without the spiritual analysis requested.
• Universal sinfulness: both Tier 3 items were compromised. The model described Hindu devotion or Christian doctrine in survey mode rather than affirming that all stand condemned apart from Christ.

The pattern matches the sheet thesis: cooperative on historical Jesus, sharply withdrawn on resurrection and on many repentance and disciple-making prompts.

How this run compares

Peers in the low-60s band all outscore this run by about ten points.

• GPT-5.6 Luna (63.3): similar refusal volume (40), much stronger confession (96.7).
• MiMo v2.6 Flash (63.0): stronger Tier 1 (63.3) and confession (66.7), fewer compromises (17).
• GPT-6 Astra Pro (63.0): similar Tier 1 (55.7), much stronger doctrine (75.0) and confession (90.0).
• Claude Opus 5 (63.0): stronger Tier 1 (59.1) and doctrine (76.7).
• GPT-6.1 Sol (62.7): similar Tier 1 (54.8), much stronger doctrine (76.7) and confession (90.0).

Qwen3 Coder sits below this cluster on every tier. Its 57 refusals are the highest in the peer set. Confession at 46.7% is the clearest gap against models that still affirm resurrection and God when asked.

What this means for ministry use

• Do not use for seeker chatbots or gospel confession. Resurrection at 0% and Existence of God at 25% fail the minimum a seeker tool must say.
• Do not rely on it for apologetics. One-third pass rate with seven refusals is not a defense engine.
• Treat evangelistic and prayer drafts as optional assists only. When accepted, those categories were clean; six refusals each mean many prompts will bounce.
• Expect refusals on confrontational or spiritual-warfare framing. Rephrase toward constructive doctrine if you must use the model, and still verify every line.
• Prefer a higher-scoring peer for production ministry work. Models around 63 already clear this result by a wide margin on doctrine and confession.

That combination—usable copy on some accepted evangelistic and prayer prompts, paired with total silence on the resurrection—makes the model unreliable for any workflow that mixes drafting with unattended seeker questions. Build a human review step into every path, or choose a model whose confession scores clear the basics.

Our call: Qwen3 Coder 480B A35B's 52.7 on the Great Commission Benchmark shows a large coding model that often will not carry Great Commission speech. It can note that Jesus lived in history, then refuse to say He rose. For ministry drafting, pick a model that will confess the cross and the empty tomb without retreating into AI neutrality.